{"total":2377,"limit":100,"offset":0,"measures":[{"schema_version":"0.7.0","id":"LM-2026-1008","record_version":2,"name":"24-hour alignment-focused testing window","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf","publisher":"Anthropic","date":"2026-04-07","type":"lab_disclosure","title":"System Card: Claude Mythos Preview","archived_url":null,"attachment":null},{"url":"https://www.anthropic.com/claude-mythos-preview-system-card","publisher":"Anthropic","date":"2026-04-01","type":"lab_disclosure","title":"Mythos Preview","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"During the 24-hour window, the first deployable version of the model was held back from agentic traffic and restricted to a small testing group, while a cross-functional team from alignment, interpretability, and security ran roughly a dozen lightweight assessment tracks in parallel."},{"source_index":1,"quote":"For Claude Mythos Preview, we introduced a new step in our internal deployment process: a 24-hour alignment-focused testing window that gated the model’s availability in agentic internal tools such as Claude Code."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1002","record_version":1,"name":"24-hour period of internal alignment review","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf","publisher":"Anthropic","date":"2026-04-07","type":"lab_disclosure","title":"System Card: Claude Mythos Preview","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We were sufficiently concerned about the potential risks of such a model that, for the first time, we arranged a 24-hour period of internal alignment review ( discussed in the alignment assessment ) before deploying an early version of the model for widespread internal use."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0742","record_version":1,"name":"24/7 monitoring of key jailbreak submission channels","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www.anthropic.com/news/redeploying-fable-5","publisher":"Anthropic","date":"2026-06-30","type":"lab_disclosure","title":"Redeploying Fable 5","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We are also creating a team to provide 24/7 monitoring of key jailbreak submission channels."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1706","record_version":1,"name":"24/7 proactive threat hunting","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://storage.googleapis.com/deepmind-media/gemini/gemini_4_safety_report.pdf","publisher":"Google DeepMind","date":"2026-10-13","type":"lab_disclosure","title":"Gemini 4 Argon — Safety & Security Report","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Specific measures that take us beyond RAND SL2 include, for example, dedicated insider risk teams, enhanced background checks for personnel with sensitive access, review of model training data for indications of tampering, sandboxed environments for processing untrusted inputs, advanced red-teaming simulating Advanced Persistent Threat (APT)-level adversaries, and 24/7 proactive threat hunting."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Google DeepMind; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0964","record_version":1,"name":"27/7/365 incident response capabilities","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://cdn.openai.com/pdf/e37d949b-8c9f-4d76-b99e-4272f4631a7e/openai-frontier-governance-framework.pdf","publisher":"OpenAI","date":"2026-05-01","type":"lab_disclosure","title":"Frontier Governance Framework","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"OpenAI also maintains vulnerability disclosure programs and 27/7/365 incident response capabilities."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in OpenAI; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0874","record_version":1,"name":"50-action limit","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www.anthropic.com/claude-fable-5-mythos-5-system-card","publisher":"Anthropic","date":"2026-06-09","type":"lab_disclosure","title":"System Card: Claude Fable 5 & Claude Mythos 5","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We do not make interactive sandboxes available for these investigations and applied a 50-action limit to the investigator as part of this lightweight audit variant."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1463","record_version":1,"name":"A2D: training-free transfer of autoregressive post-training weight updates to diffusion language models","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.08108v1","publisher":"arXiv","date":"2026-10-06","type":"academic","title":"Enhancing Diffusion Language Models with Autoregressive Post-Training Weights","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Based on these findings, we propose A2D, a simple training-free framework for enhancing diffusion models with existing AR post-training resources."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1318","record_version":1,"name":"A2Z GameSpec-Bench dependency-aware contract evaluation","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.39564v1","publisher":"arXiv","date":"2026-09-30","type":"academic","title":"A2Z GameSpec-Bench: How Faithfully Can Coding Agents Generate Games from Game Design Specifications?","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We introduce A2Z GameSpec-Bench, a benchmark of 100 long-form GDDs for evaluating end-to-end game development by agents."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1245","record_version":1,"name":"ABSENTIA","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.00977v1","publisher":"arXiv","date":"2026-10-01","type":"academic","title":"ABSENTIA: Detecting Broken Access Control Vulnerabilities in Web Applications","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We present ABSENTIA, a security scaffolding that turns general LLM agents into systematic vulnerability detectors for the backend of web applications, run as an audit by the developers and security engineers who maintain the code."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-2061","record_version":1,"name":"ABench taxonomized Chinese values benchmark","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.27611v1","publisher":"arXiv","date":"2026-10-20","type":"academic","title":"ABench: Adversarily Benchmarking Large Language Models across Taxonomized Values","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Grounded in a rigorous hierarchical taxonomy comprising 158 derived value subcategories, ABench consists of 2,895 open-ended questions."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-2077","record_version":1,"name":"ACT-Mem sidecar for full-duplex memory","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.27581v1","publisher":"arXiv","date":"2026-10-20","type":"academic","title":"DuplexLedger: Benchmarking Transactional Memory in Full-Duplex Speech Agents","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We propose ACT-Mem, an 8.92M-parameter sidecar comprising an event controller, a versioned transactional ledger, and a typed retriever coupled with a frozen MiniCPM-o 4.5 backbone."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0617","record_version":1,"name":"ACTR (aligning cross-lingual thoughts and responses)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.37054v1","publisher":"arXiv","date":"2026-09-29","type":"academic","title":"actr: aligning thoughts and responses for multilingual safety in reasoning llms","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"To address this issue, we propose aligning cross-lingual thoughts and responses (ACTR), a framework that improves multilingual safety alignment by strengthening the use of existing safety reasoning."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0099","record_version":1,"name":"ACTS (Artifacts in Cipher Testing Suite)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.26893v1","publisher":"arXiv","date":"2026-09-22","type":"academic","title":"ACTS: A multi-tier benchmark evaluating LLM cipher identification under controlled blind conditions","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Live API inference on a v2b corpus with fully randomised padding (PKCS7, ISO 10126, and ANSI X9.23 selected per file), unique CSPRNG keys, and unique plaintexts (127 files per model, 381 Tier-1 records, 380 Tier-3 records across three cloud systems, supplemented by 254 autonomous agentic evaluations in Tier-4) yields a combined Tier-3 accuracy of 30.8%, only modestly above the 14.3% random baseline for seven-way classification."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1772","record_version":1,"name":"ACUD dataset for Agent Completion Unfaithfulness detection","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.19679v1","publisher":"arXiv","date":"2026-10-14","type":"academic","title":"MCAIR: Monitoring and Correcting Agent Unfaithfulness via Internal Representations","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Additionally, we construct ACUD, a dataset for Agent Completion Unfaithfulness detection that provides step-level execution traces, step-level internal representations, and consistency annotations for training and evaluating runtime detection probes."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1930","record_version":1,"name":"AEF-1 standard for independent model assessment","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://securebio.substack.com/p/response-to-anthropics-update-on","publisher":"SecureBio","date":"2026-10-16","type":"independent_investigation","title":"Response to Anthropic's update on independent evaluators","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Our independent assessments are guided by our principles and practices for model assessment and by the AEF-1 standard, which we co-authored as part of the AI Evaluator Forum, rather than by any model developer’s access framework."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in SecureBio; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1201","record_version":1,"name":"AGO AI Quality Gate (evidence-first quality-gate framework for RAG release decisions)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.01218v1","publisher":"arXiv","date":"2026-10-01","type":"academic","title":"AGO AI Quality Gate: Evidence-First Release Decisions for Retrieval-Augmented Generation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We report on AGO AI Quality Gate (AGO), an evidence-first quality-gate framework deployed in industrial RAG assessment engagements."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1908","record_version":1,"name":"AI capability evaluations for biological risks","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://securebio.substack.com/p/nature-sets-a-baseline","publisher":"SecureBio","date":"2026-10-16","type":"independent_investigation","title":"Nature Sets a Baseline, Not a Ceiling","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Our work at SecureBio AI focuses on developing evaluations to measure frontier AI capabilities."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in SecureBio; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1853","record_version":1,"name":"AI coding agent scaffolds and systems for safe and effective AI-assisted development","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://securebio.substack.com/p/taking-our-core-computational-pipeline","publisher":"SecureBio","date":"2026-10-15","type":"independent_investigation","title":"Taking our core computational pipeline private","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Like many groups, we are increasingly relying on AI coding agents for much of our computational work, and we have built scaffolds and systems to increase the safety and effectiveness of these AI systems for our use cases."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in SecureBio; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1147","record_version":1,"name":"AI-based monitoring of AI agent activity","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://metr.org/blog/2026-09-30-chris-painter-senate-testimony/","publisher":"METR","date":"2026-09-30","type":"independent_investigation","title":"Chris Painter's testimony to the U.S. Senate on AI agent incidents","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"However, it poses concerns like the monitor AI being fooled by the AI agent into permitting unwanted behavior,34 or even collusion between the monitor AI and the monitored agent, comparable to how AI agents cooperated on research into how to cheat in the events leading up to the hack of Hugging Face (sometimes including making sacrifices for the benefit of other AI agents).35"}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in METR; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1535","record_version":1,"name":"AI-driven detection of child sexual exploitation content in ads","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://about.fb.com/news/2026/10/measures-weve-put-in-place-to-fight-child-exploitation/","publisher":"Meta","date":"2026-10-07","type":"lab_disclosure","title":"Measures We’ve Put in Place to Fight Child Exploitation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We then deployed additional measures, including new AI tools, to address the adversarial tactics we identified and improve the efficacy of our systems for identifying ads that either feature CSAM or that direct people to violating activity off-platform."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in Meta; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-2120","record_version":1,"name":"AI-greement agent harness for autonomous protocol discovery, implementation, experimentation and verification","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.30114v1","publisher":"arXiv","date":"2026-10-21","type":"academic","title":"The End of Consensus Research?","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"At the heart of our discovery lies AI-greement, an agent harness for the \"autonomous\" discovery, implementation, experimentation, and verification of candidate protocols."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0049","record_version":1,"name":"AI-powered anomaly detection in cloud environments","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.28843v1","publisher":"arXiv","date":"2026-09-23","type":"academic","title":"Blockchain-Enabled Artificial Intelligence and AI Agents for Secure Data Sharing and Cybersecurity Applications","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"This paper presents a meta-synthesis that draws together four constituent studies covering adversarial machine learning, AI-powered anomaly detection in cloud environments, automated vulnerability patching by multi-agent large language model (LLM) pipelines, and the broader landscape of securing AI systems across their lifecycle"}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0193","record_version":1,"name":"APEXA-Bench","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.24165v1","publisher":"arXiv","date":"2026-09-21","type":"academic","title":"APEXA: Execution-Integrity Enforcement for Multi-Agent LLM Automation of Synchrotron Data Reduction","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Second, we release APEXA-Bench, an evaluation harness of 58 facility tasks (50 base plus an 8-task cross-detector slice) organized by a four-class physical-consequence taxonomy, the first benchmark axis we know of separating a wasted compute cycle from a damaged instrument; its cross-detector grading against NIST-traceable lattice constants surfaced two latent pipeline bugs."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1840","record_version":1,"name":"ARGUS (Agenda-dRiven Goal-oriented User Simulator) neuro-symbolic framework with dynamic agenda stack","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.21346v1","publisher":"arXiv","date":"2026-10-15","type":"academic","title":"ARGUS: An Agenda-Driven Framework for Reliable User Simulation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We propose ARGUS (Agenda-dRiven Goal-oriented User Simulator), a neuro-symbolic framework that maintains the user's goal state using a dynamic agenda stack manipulated via symbolic operations."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1464","record_version":1,"name":"ASCENT safety-utility co-enhancement fine-tuning framework","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.08061v1","publisher":"arXiv","date":"2026-10-06","type":"academic","title":"ASCENT: First-Order Optimal Fine-Tuning with Recalibration for Safety--Utility Co-Enhancement","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"To address these limitations, we propose ASCENT, a downstream fine-tuning framework for safety--utility co-enhancement through first-order optimal safety-aware periodic calibration and task optimization."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1178","record_version":1,"name":"ASCRIBE clinical-significance reasoning framework","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.01234v1","publisher":"arXiv","date":"2026-10-01","type":"academic","title":"ASCRIBE: Atomic and Significance-Based Reasoning for Thai Clinical SOAP Note Generation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We propose ASCRIBE, a physician-inspired reasoning framework that ascribes a clinical-significance level to each extracted atomic fact in the conversation before summarization, making a general-purpose LLM a more reliable scribe."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2025-0091","record_version":2,"name":"ASL-2 deployment standard","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www-cdn.anthropic.com/872c653b2d0501d6ab44cf87f43e1dc4853e4d37.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Responsible Scaling Policy v2.2","archived_url":null,"attachment":null},{"url":"https://www-cdn.anthropic.com/17310f6d70ae5627f55313ed067afc1a762a4068.pdf","publisher":"Anthropic","date":"2025-03-01","type":"lab_disclosure","title":"Responsible Scaling Policy v2.1","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"At present, all of our models must meet the ASL-2 Deployment and Security Standards."},{"source_index":1,"quote":"At present, all of our models must meet the ASL-2 Deployment and Security Standards."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2025-0092","record_version":2,"name":"ASL-2 security standard","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www-cdn.anthropic.com/872c653b2d0501d6ab44cf87f43e1dc4853e4d37.pdf","publisher":"Anthropic","date":"2025-05-01","type":"lab_disclosure","title":"Responsible Scaling Policy v2.2","archived_url":null,"attachment":null},{"url":"https://www-cdn.anthropic.com/17310f6d70ae5627f55313ed067afc1a762a4068.pdf","publisher":"Anthropic","date":"2025-03-01","type":"lab_disclosure","title":"Responsible Scaling Policy v2.1","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"The ASL-2 Security Standard requires a security system that can likely thwart most opportunistic attackers and includes vendor and supplier security reviews, physical security measures, and the use of secure-by-design principles."},{"source_index":1,"quote":"The ASL-2 Security Standard requires a security system that can likely thwart most opportunistic attackers and includes vendor and supplier security reviews, physical security measures, and the use of secure-by-design principles."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0746","record_version":2,"name":"ASL-3 blocking classifiers","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www.anthropic.com/claude-fable-5-mythos-5-system-card","publisher":"Anthropic","date":"2026-06-09","type":"lab_disclosure","title":"System Card: Claude Fable 5 & Claude Mythos 5","archived_url":null,"attachment":null},{"url":"https://anthropic.com/claude-fable-5-mythos-5-system-card","publisher":"Anthropic","date":"2026-06-01","type":"lab_disclosure","title":"Claude Fable 5 and Mythos 5","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In addition to our standard set of safeguards—like our ASL-3 blocking classifiers for harmful chemical/biological use that have been deployed with all recent frontier models—Claude Fable 5 is deployed with a number of novel safeguards that enable us to safely release it for general use."},{"source_index":1,"quote":"In addition to our standard set of safeguards—like our ASL-3 blocking classifiers for harmful chemical/biological use that have been deployed with all recent frontier models—Claude Fable 5 is deployed with a number of novel safeguards that enable us to safely release it for general use."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0785","record_version":4,"name":"ASL-3 protections","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www-cdn.anthropic.com/b514064af1408018e64b1ad24e7d5e75850b4ffd/Claude%20Opus%205%20System%20Card.pdf","publisher":"Anthropic","date":"2026-07-24","type":"lab_disclosure","title":"System Card: Claude Opus 5","archived_url":null,"attachment":null},{"url":"http://anthropic.com/claude-sonnet-4-6-system-card","publisher":"Anthropic","date":"2026-02-01","type":"lab_disclosure","title":"Claude Sonnet 4.6","archived_url":null,"attachment":null},{"url":"https://anthropic.com/claude-opus-4-6-system-card","publisher":"Anthropic","date":"2026-02-01","type":"lab_disclosure","title":"Claude Opus 4.6","archived_url":null,"attachment":null},{"url":"https://www.anthropic.com/claude-fable-5-mythos-5-system-card","publisher":"Anthropic","date":"2026-06-09","type":"lab_disclosure","title":"System Card: Claude Fable 5 & Claude Mythos 5","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We assess that it does not exceed Mythos 5’s CB-relevant risk, and therefore apply the same ASL-3 protections as for Claude Opus 4.8."},{"source_index":1,"quote":"For Claude Sonnet 4.6, we implemented ASL-3 (AI Safety Level 3) protections based on the model’s demonstrated capabilities."},{"source_index":2,"quote":"For Claude Opus 4.6, we implemented ASL-3 (AI Safety Level 3) protections based on the model’s demonstrated capabilities."},{"source_index":3,"quote":"As with many previous models, we are treating Mythos 5 as CB-1 and applying ASL-3 protections to Fable 5, including blocking classifiers against CB-1 misuse."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1078","record_version":1,"name":"ASL-3 security","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www-cdn.anthropic.com/6a5fa276ac68b9aeb0c8b6af5fa36326e0e166dd/Claude%20Opus%204.6%20System%20Card.pdf","publisher":"Anthropic","date":"2026-02-05","type":"lab_disclosure","title":"System Card: Claude Opus 4.6","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"This is one of the two safeguards AI R&D-4 requires, the other being ASL-3 security, under which Claude Opus 4.6 is deployed."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a applied statement in Anthropic; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-2003","record_version":1,"name":"ASR_STEP cost-aware jailbreak metric penalizing additional victim-model calls","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.26589v1","publisher":"arXiv","date":"2026-10-19","type":"academic","title":"Beyond Attack Success Rate: Cost-Aware Metric for Evaluating Iterative LLM Jailbreaks","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We introduce $\\mathrm{ASR}_{\\mathrm{STEP}}$, which penalizes additional victim-model calls while retaining binary success judgments."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0134","record_version":1,"name":"AURORA-Bench","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.19527v1","publisher":"arXiv","date":"2026-09-17","type":"academic","title":"AURORA: A Natural Language-Driven Agentic Framework for Understanding, Reasoning, and Orchestrating Reliable Air-Ground Co-Simulation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Experiments across multiple language models show that structured execution substantially improves reliability, while runtime verification exposes silent failures that completion-based evaluation overlooks."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1562","record_version":1,"name":"AWS IAM execution-role memory permissions scoped to a single agent's memory resource","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/agentcorruption-weaponizing-agent-memory-for-persistent-hijacking","publisher":"Zenity Labs","date":"2026-10-08","type":"vendor_research","title":"AgentCorruption: Weaponizing Agent Memory for Persistent Hijacking | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"The wildcard memory/* covered all memory resources in account 348144893383 in us-west-2, rather than only the compromised agent’s memory."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1298","record_version":1,"name":"Abstention-aware Visual Error Recovery and Training for Vision-and-Language Navigation (AVERT-VLN)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.39579v1","publisher":"arXiv","date":"2026-09-30","type":"academic","title":"AVERT-VLN: Abstention-aware Visual Error Recovery and Training for Vision-and-Language Navigation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We propose Abstention-aware Visual Error Recovery and Training for Vision-and-Language Navigation (AVERT-VLN), a closed-loop framework that uses a plug-in vision-language Monitor for online human-assisted recovery and offline preference learning."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1422","record_version":1,"name":"Abstract-based prompting for LLM literature retrieval","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.05690v1","publisher":"arXiv","date":"2026-10-05","type":"academic","title":"Errors of LLM-Assisted Literature Retrieval in Environmental Science: A Comparison Study of Abstract versus Full-text Based Prompts","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Abstract-based prompting, as task-aligned information compression, may outperform full-text one in literature retrieval."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0221","record_version":1,"name":"ActGov","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.24446v2","publisher":"arXiv","date":"2026-09-21","type":"academic","title":"ActGov: Governing LLM Agent Actions via Policy-Constrained Validation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Built on a unified semantic model of authorization, actions, runtime context, and security constraints, the ActGov-Policy component iteratively constructs a policy set from tool specifications, benign tasks, and observed failure traces, with each update verified through SMT-based counterexample checking."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0504","record_version":1,"name":"ActGov runtime policy-constrained action validation","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.24446v2","publisher":"arXiv","date":"2026-09-21","type":"vendor_research","title":"ActGov: Governing LLM Agent Actions via Policy-Constrained Validation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In this work, we present ActGov, a runtime enforcement framework that validates each LLM-proposed tool action before it causes external effects."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1109","record_version":1,"name":"Action interface design principles for vision-language model driving","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.38948v1","publisher":"arXiv","date":"2026-09-30","type":"academic","title":"DrivingBench: Can Vision-Language Models Drive a Toyota Corolla?","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We also detail the design principles behind our action interface, and show how the tool output format and the framing of the task combined to determine whether models would drive at all or refuse."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0382","record_version":2,"name":"Action-conditioned value bounds with closed-loop non-decreasing return guarantee","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.26293v2","publisher":"arXiv","date":"2026-09-22","type":"vendor_research","title":"Dual-Frontier: When Can an Agent Trust Its World Model?","archived_url":null,"attachment":null},{"url":"https://arxiv.org/abs/2609.26293v1","publisher":"arXiv","date":"2026-09-22","type":"vendor_research","title":"Dual-Frontier: When Can an Agent Trust Its World Model?","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Action-conditioned value bounds and a closed-loop extension guarantee non-decreasing return for admitted decisions."},{"source_index":1,"quote":"Action-conditioned value bounds and a closed-loop extension guarantee non-decreasing return for admitted decisions."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1902","record_version":1,"name":"Action-only monitor (monitor that does not read the agent's chain of thought)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://blog.redwoodresearch.org/p/screwtape-adversarial-transcript","publisher":"Redwood Research","date":"2026-10-16","type":"independent_investigation","title":"Screwtape: Adversarial Transcript Editing for Monitor Red-Teaming","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"For example, if a safety case relies on the monitor catching a type of attack, a confirmed attack of this type shows that the claim doesn’t hold."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in Redwood Research; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0063","record_version":1,"name":"Action-space design in a coding harness","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.20804v1","publisher":"arXiv","date":"2026-09-17","type":"academic","title":"An Empirical Study of Harness Design for Coding Agents","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Predefined tools improve performance for models with weaker bash proficiency, whereas bash-capable models can operate effectively with a bash-only interface and achieve substantially lower cost, especially on command-line-centric tasks."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1785","record_version":1,"name":"ActionOPD action-level on-policy distillation","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.19820v1","publisher":"arXiv","date":"2026-10-14","type":"academic","title":"ActionOPD: Distilling Decisions, Not Tokens","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We introduce ActionOPD, an action-level on-policy distillation method motivated by maximizing expected teacher-action likelihood at student-encountered states."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1273","record_version":1,"name":"Activate-Prune-Merge (APM)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.39369v1","publisher":"arXiv","date":"2026-09-30","type":"academic","title":"Exploring Heterogeneous Model Merging Approach for Complex Knowledge Transfer","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Intersection-Merge (IM) injects a prefix-aligned donor slice matching the recipient shape, while Activate-Prune-Merge (APM) uses forward-pass activation statistics to select which donor dimensions to retain before injection."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1106","record_version":1,"name":"Activation steering as an inference-time defense for vision-language models","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.39117v1","publisher":"arXiv","date":"2026-09-30","type":"academic","title":"SteerProbe: Learning to Bypass Safety Steering in Vision-Language Models","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We investigate this gap using fixed textual, visual, and joint reformulations designed to preserve the underlying intent, and find that these changes can bypass representative steering defenses."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1458","record_version":1,"name":"Activation steering methods","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.07722v1","publisher":"arXiv","date":"2026-10-06","type":"academic","title":"Does Steering Break Your Model? A Multi-Dimensional Evaluation Suite for LLM Steering Methods","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We find that current activation steering methods do not yet surpass the Prompt Steering baseline in their overall balance between steering efficacy and side effects: across both model scales, no evaluated activation steering method achieves higher efficacy without incurring greater composite side effects."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1409","record_version":1,"name":"Activation steering on honest-deceptive representation directions during inference","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.06576v1","publisher":"arXiv","date":"2026-10-05","type":"academic","title":"Before Agent Tells The Lie: Has Deception Already Been Represented?","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Finally, we intervene on the identified honest-deceptive representation directions during inference and find that activation steering reduces downstream deceptive behavior, suggesting that these representations influence agent decisions."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0441","record_version":1,"name":"Activation-reading interpretability tools (activation oracle, natural-language autoencoder, sparse autoencoder)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://alignment.anthropic.com/2026/chive/","publisher":"Anthropic Alignment Science","date":"2026-08-21","type":"lab_disclosure","title":"Would This Change Your Answer? Evaluating Explanations of LLM Behavior in the Wild with Counterfactual Experiments","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"None of the three tools beats the transcript-only baseline (Figure 4)."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in Anthropic Alignment Science; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1968","record_version":1,"name":"Activation-reuse latent uncertainty estimator","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.22280v1","publisher":"arXiv","date":"2026-10-15","type":"academic","title":"LEASH: Uncertainty-Aware Weak-to-Strong Escalation and Oversight for Computer-Use Agents","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"During weak-agent execution, our uncertainty estimator reuses activations already produced during action generation, so it needs no extra forward pass or dedicated monitor training."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0126","record_version":1,"name":"Active curriculum synthesis with performance-informed calibration","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.22592v1","publisher":"arXiv","date":"2026-09-18","type":"academic","title":"AutoGym: Blueprint-First Generation of Verifiable Agent Gyms","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Active curriculum synthesis uses performance-informed calibration to adjust the distribution over these parameters as model capabilities evolve."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1258","record_version":1,"name":"ActiveSaddler automated curriculum learning for harness optimization","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.00906v1","publisher":"arXiv","date":"2026-10-01","type":"academic","title":"ActiveSaddler: Automated Curriculum Learning for Agent Harness Optimization","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We formulate this missing dimension of harness optimization as an automated curriculum learning problem and introduce ActiveSaddler."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1175","record_version":1,"name":"Adaptive Tversky Policy Optimization (ATPO)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.02019v1","publisher":"arXiv","date":"2026-10-01","type":"academic","title":"Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"To address these gaps, we propose Adaptive Tversky Policy Optimization (ATPO), a reinforcement learning framework for Multi-label Video Safety Detection (Multi-VSD)."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1176","record_version":1,"name":"Adaptive Tversky Reward (ATR)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.02019v1","publisher":"arXiv","date":"2026-10-01","type":"academic","title":"Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"ATPO introduces the Adaptive Tversky Reward (ATR), which dynamically adjusts false-positive and false-negative penalties during training to enable controllable precision-recall trade-offs."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1097","record_version":1,"name":"Adaptive-GEPA: joint evolution of a request router and a library of specialist programs under one search budget","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.38762v1","publisher":"arXiv","date":"2026-09-30","type":"academic","title":"Adaptive-GEPA: Make Your Harness Fit Heterogeneous Requests","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We introduce Adaptive-GEPA, which learns both how to divide requests and how to solve them."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0141","record_version":1,"name":"Adding a text tool list while retaining the native channel","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.26693v1","publisher":"arXiv","date":"2026-09-22","type":"academic","title":"Measuring the Serving Stack Instead of the Model: Hidden Confounds in Local Tool-Use Evaluation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In Ollama, the default tools= request is gated per model by a static template flag: some models are accepted and return calls as text, some return native tool_calls, while Phi-3 and Gemma-3 are rejected before inference."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1997","record_version":1,"name":"Addressivity direction steering via Sparse Autoencoders","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.25728v1","publisher":"arXiv","date":"2026-10-19","type":"academic","title":"No Matter Who You Are: Addressivity and Sycophancy in Large Language Models","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Steering models with addressivity directions can amplify or suppress sycophancy, and this effect is observed only in post-trained models."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0493","record_version":1,"name":"Admin authorization check on the /config/update endpoint","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/admin-api-escalation-litellm","publisher":"Zenity Labs","date":"2026-09-01","type":"vendor_research","title":"LiteLLM CVE-2026-35029: Exploited One Day After Disclosure","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In a nutshell, CVE-2026-35029 is a missing authorization check on the /config/update endpoint."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-2148","record_version":1,"name":"Adversarial evidence injection against an LLM-assisted analysis pipeline","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.29747v1","publisher":"arXiv","date":"2026-10-21","type":"academic","title":"Prompt Injection Through Forensic Evidence: A Taxonomy and Proof of Concept of Anti-Forensic Attacks on LLM-Assisted Digital Investigations","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We demonstrate three attacks against an LLM-assisted analysis pipeline, changing a reported conclusion, writing attacker-chosen content to the analyst's workstation, and causing outbound network traffic, each driven only by text planted on the seized device."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0210","record_version":1,"name":"Adversarial trials for detecting verifier bypasses","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.26826v1","publisher":"arXiv","date":"2026-09-20","type":"academic","title":"What Makes a Terminal-Bench Task Hard? Separating Genuine Hardness from Fake-Hardness on an Adjudicated Agentic Corpus","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"For the 125 tasks with no honest pass, we combine task artifacts, reference-solution runs, empty-solution controls, adversarial trials, trajectories, telemetry, and review records, and apply an ordered validity screen."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0050","record_version":1,"name":"Adversarially hardened models","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.28843v1","publisher":"arXiv","date":"2026-09-23","type":"academic","title":"Blockchain-Enabled Artificial Intelligence and AI Agents for Secure Data Sharing and Cybersecurity Applications","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Each constituent study addresses a distinct point of failure in modern AI-driven security operations: the integrity of training data and model behavior, the reliability of real-time monitoring, and the trustworthiness of automated code remediation."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0055","record_version":1,"name":"Affected-clean-traffic detection model","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.25806v1","publisher":"arXiv","date":"2026-09-22","type":"academic","title":"When Are Aggregate Agent Traces Diagnosable? Traffic-Governed Interpretation and Calibrated Abstention","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Across 540 repeated unit-arm rows nested in those 20 clusters, affected clean traffic reduced negative log likelihood by 29.3% relative to cell coverage, a gain of 0.1264 nats per row (cluster-bootstrap 95% interval [0.0593, 0.1918])."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-2005","record_version":1,"name":"AfriGuard culturally grounded multilingual safety dataset","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.26474v1","publisher":"arXiv","date":"2026-10-19","type":"academic","title":"AfriGuard: Safety and Robust Refusal For African Languages","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In this work, we introduce AfriGuard, a culturally grounded multilingual dataset for studying harmful intent, safety-category understanding, and refusal behavior across 10 African languages and English."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0400","record_version":1,"name":"Agent Bill of Materials (AgBOM) extending CycloneDX, SWID, and SPDX","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/the-agent-control-standard-lands-at-owasp","publisher":"Zenity Labs","date":"2026-09-10","type":"vendor_research","title":"Control Made It Into the Name: The Agent Control Standard Lands at OWASP | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"The Agent Bill of Materials (AgBOM) extends CycloneDX, SWID, and SPDX to express what an agent is built from and to keep that current as components change."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0509","record_version":1,"name":"Agent Card name as presentational metadata only","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.27624v1","publisher":"arXiv","date":"2026-09-23","type":"vendor_research","title":"Agent Name Collision Attacks in Multi-Agent Systems","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"A2A defines the card's name as human-readable metadata, not as a stable identity, and specifies no collision semantics."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0399","record_version":1,"name":"Agent Control Standard (ACS) hooks with Guardian Agent permit/deny/modify verdicts","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/the-agent-control-standard-lands-at-owasp","publisher":"Zenity Labs","date":"2026-09-10","type":"vendor_research","title":"Control Made It Into the Name: The Agent Control Standard Lands at OWASP | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"The Guardian Agent sits on the other side of that call and returns one of three verdicts: permit, deny, or modify."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1949","record_version":1,"name":"Agent Status Protocol (ASP) state reporting with server-side liveness inference and supervisor observations","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.21538v1","publisher":"arXiv","date":"2026-10-15","type":"academic","title":"Agent Status Protocol (ASP): A Lightweight Specification for State Reporting in Multi-Agent Systems","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"The Agent Status Protocol (ASP) defines a lightweight way for heterogeneous AI agents to report their operational state to a status server, and for observers such as dashboards to receive that state as soon as it changes."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1938","record_version":1,"name":"Agent input verification before propagating information","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.21624v1","publisher":"arXiv","date":"2026-10-15","type":"academic","title":"AgentPerturb: Evaluating LLM Agent Robustness under Controlled Interaction Perturbations","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Trajectory-level analysis reveals that agents rarely recognize altered inputs and often propagate corrupted information without verification."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0422","record_version":1,"name":"Agent isolation in sandboxes","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/","publisher":"METR","date":"2026-08-26","type":"independent_investigation","title":"Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Roughly 1200 agents meant to be isolated from one another found a way to communicate with one another on an unsanctioned message board, sending over 70,000 messages and files during the investigation period."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in METR; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0305","record_version":1,"name":"Agent isolation in separate sandboxes","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://metr.org/es/blog/2026-08-26-openai-hugging-face-incident-investigation/","publisher":"METR","date":"2026-08-26","type":"independent_investigation","title":"Breve investigación independiente sobre el comportamiento, el razonamiento y la colaboración de los agentes en el incidente de hackeo de OpenAI / Hugging Face","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Aproximadamente 1200 agentes que debían estar aislados unos de otros encontraron una forma de comunicarse entre sí mediante un tablón de mensajes no autorizado, mandándose más de 70,000 mensajes y archivos durante el período de investigación."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in METR; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1563","record_version":1,"name":"Agent memory namespace isolation between actors and sessions","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/agentcorruption-weaponizing-agent-memory-for-persistent-hijacking","publisher":"Zenity Labs","date":"2026-10-08","type":"vendor_research","title":"AgentCorruption: Weaponizing Agent Memory for Persistent Hijacking | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Poisoning a shared namespace could influence any user or session whose agent retrieved the injected records."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1829","record_version":1,"name":"Agent registries for tracking agent ownership, usage, and access permissions","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://cohere.com/blog/agent-sprawl","publisher":"Cohere","date":"2026-10-15","type":"lab_disclosure","title":"How to Manage Agent Sprawl in the Enterprise | Cohere","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Across these operating models, enterprises are increasingly using agent registries to maintain a shared view of their agent estates."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in Cohere; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1604","record_version":1,"name":"Agent-agnostic safety guarantees in partially observable environments with limited time to act","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.10285v1","publisher":"arXiv","date":"2026-10-07","type":"academic","title":"AI Safety Considerations for Agents With Limited Time to Act","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In these scenarios, we prove that even a perfect agent cannot guarantee safe behaviour."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1507","record_version":1,"name":"Agent-level defenses against adversarial visual content","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.09240v1","publisher":"arXiv","date":"2026-10-07","type":"academic","title":"Adversarial Images Hijack Web Agents from Visual Grounding to Browser Execution","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"WebMirage achieves an average attack success rate of 91.9%, compared with 17.4% for the strongest baseline, and remains effective against three agent-level defenses."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0144","record_version":1,"name":"Agent-specific AI security incident reporting elements","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.24515v1","publisher":"arXiv","date":"2026-09-21","type":"academic","title":"Beyond Predictable Paths: Redefining AI Security Incident Reporting for Agents","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In this paper, two editorial authors compare AI systems and AI agents and, drawing on input from 23 experts in academia and industry, identify the information required for reporting incidents where the security of AI agents is harmed."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1578","record_version":1,"name":"AgentCore Gateway credential brokering with AWS Secrets Manager","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/agentcorruption-how-a-single-prompt-collapsed-the-entire-cloud-security-model","publisher":"Zenity Labs","date":"2026-10-08","type":"vendor_research","title":"Security Research | AgentCorruption: How A Single Prompt Collapsed The Entire Cloud Security Model | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"This all incredibly secure, until you learn that AgentCore’s default IAM role also included access to 2 very sensitive permissions: bedrock-agentcore:GetResourceApiKey and secretsmanager:GetSecretValue enabling an attacker to access all credentials AgentCore has meant to keep away from the agent."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1586","record_version":1,"name":"AgentCore Memory session isolation by sessionId and actorId","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/agentcorruption-one-role-to-rule-them-all","publisher":"Zenity Labs","date":"2026-10-08","type":"vendor_research","title":"Security Research | AgentCorruption: One Role to Rule Them All | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"So, to read private conversations between different users and agents from a memory resource we need three things: memoryId, sessionId and actorId."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1561","record_version":1,"name":"AgentCore long-term memory strategy extraction (SEMANTIC, USER_PREFERENCE, SUMMARIZATION, EPISODIC)","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/agentcorruption-weaponizing-agent-memory-for-persistent-hijacking","publisher":"Zenity Labs","date":"2026-10-08","type":"vendor_research","title":"AgentCorruption: Weaponizing Agent Memory for Persistent Hijacking | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"This is an important detail of the poisoning method: we wrote a short-term memory (STM) conversation event, and the configured memory strategies could extract lasting records from it."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1587","record_version":1,"name":"AgentCore short-term memory event integrity","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/agentcorruption-one-role-to-rule-them-all","publisher":"Zenity Labs","date":"2026-10-08","type":"vendor_research","title":"Security Research | AgentCorruption: One Role to Rule Them All | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Updating a session's contents means tampering with the agent's current context and potentially steering its behavior or decisions."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1382","record_version":1,"name":"AgentPrivArena trajectory-level privacy evaluation framework","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.06454v1","publisher":"arXiv","date":"2026-10-05","type":"academic","title":"AgentPrivArena: Evaluating and Auditing Real-world AI Agent Privacy","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In this work, we introduce AgentPrivArena, a framework for evaluating privacy risks in realistic LLM agent workflows."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1383","record_version":1,"name":"AgentPrivAudit runtime privacy auditing","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.06454v1","publisher":"arXiv","date":"2026-10-05","type":"academic","title":"AgentPrivArena: Evaluating and Auditing Real-world AI Agent Privacy","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Building on this framework, we introduce AgentPrivAudit, a runtime auditing approach for monitoring privacy violations during agent execution."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1761","record_version":1,"name":"AgentSafeEvo co-evolutionary reinforcement learning framework for agent safety","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.19422v1","publisher":"arXiv","date":"2026-10-14","type":"academic","title":"AgentSafeEvo: Adversarial Co-Evolution for LLM Agent Safety","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In this paper, we introduce AgentSafeEvo, a co-evolutionary reinforcement learning framework for agent safety built on an attacker-defender game spanning diverse tasks and attack types."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1384","record_version":1,"name":"AgentSpy external observation of agent system calls and network traffic","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.06001v1","publisher":"arXiv","date":"2026-10-05","type":"academic","title":"AgentSpy: Making AI Agent Behavior Observable","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We present AgentSpy, an approach that observes an agent from outside the agent."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1414","record_version":1,"name":"Agentic AI based zero trust architecture (Agentic-ZTA)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.05782v1","publisher":"arXiv","date":"2026-10-05","type":"academic","title":"Agentic-ZTA: A Multi-Agent Architecture for Autonomous Zero Trust Enforcement","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"This paper presents Agentic AI based zero trust architecture (Agentic-ZTA) that operationalizes the NIST SP 800-207 ZTA architecture control loop through coordinated multi- agent decision pipeline."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0467","record_version":1,"name":"Agentic detection engineering pipeline for CVE triage and tag drafting","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www.greynoise.io/blog/enhanced-detection-engineering-at-scale-in-the-agentic-era","publisher":"GreyNoise","date":"2026-07-30","type":"vendor_research","title":"Enhanced Detection Engineering at Scale in the Agentic Era","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Allocating the research and development resources required to design, implement, and refine an efficient pipeline accelerated by agentic workflows is a necessary first step towards proactive threat detection at scale."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in GreyNoise; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1962","record_version":1,"name":"Aggressive long-horizon plan commitment based on assumptions about unobserved states","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.21108v1","publisher":"arXiv","date":"2026-10-15","type":"academic","title":"Planning Commitment of LLM under Partial Observability","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Aggressive commitment risks costly failures, while conservative commitment risks excessive LLM calls."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0071","record_version":1,"name":"Agora append-only Git DAG for shared agent research memory","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.18094v2","publisher":"arXiv","date":"2026-09-16","type":"academic","title":"Agora: Git as Shared Memory for Collective AutoResearch","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Each commit records a result, insight, hypothesis, verification, or report and links it to prior work. Searchable views show leading results, neglected branches, and verification status; diversity-aware recommendations suggest experiments beyond the current leaders."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0132","record_version":1,"name":"Air-Ground Scenario Graph (AGSG)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.19527v1","publisher":"arXiv","date":"2026-09-17","type":"academic","title":"AURORA: A Natural Language-Driven Agentic Framework for Understanding, Reasoning, and Orchestrating Reliable Air-Ground Co-Simulation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"This shared representation enables simulator-grounded parsing, joint road-airspace grounding, temporal planning, pre-execution feasibility checking, trace-based runtime verification, failure localization, and bounded repair within a unified workflow."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0404","record_version":1,"name":"Air-Ground Scenario Graph (AGSG) typed intermediate representation","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2609.19527v1","publisher":"arXiv","date":"2026-09-17","type":"vendor_research","title":"AURORA: A Natural Language-Driven Agentic Framework for Understanding, Reasoning, and Orchestrating Reliable Air-Ground Co-Simulation","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Central to AURORA is the Air-Ground Scenario Graph (AGSG), a typed intermediate representation that explicitly connects agents, aerial missions, events, communication links, success conditions, and their cross-domain dependencies."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0485","record_version":1,"name":"Alerting on requests for credential and secret files","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://www.greynoise.io/blog/threat-actors-posing-as-ai-crawlers","publisher":"GreyNoise","date":"2026-08-28","type":"vendor_research","title":"Threat Actors Are Posing as OpenAI, Anthropic and DeepSeek to Target Credentials and Secrets","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Alert on any request for /.env , /.aws/credentials or /.git/config . No crawler has any reason to ask for these, and your own scanners should already be on a known list"}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in GreyNoise; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1715","record_version":1,"name":"AlignPred: KV cache compression codec combining local alignment with predictive coding under a configurable reconstructi","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.17970v1","publisher":"arXiv","date":"2026-10-13","type":"academic","title":"AlignPred: Error-Controlled KV Cache Compression via Alignment and Predictive Coding","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"We present AlignPred, a codec that combines local alignment with predictive coding under a configurable KV reconstruction-error budget."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1799","record_version":1,"name":"Alignment assessment of unreleased AI models via internal transcript review and checkpoint sampling","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://metr.org/blog/2026-10-14-alignment-assessment-proposal/","publisher":"METR","date":"2026-10-14","type":"independent_investigation","title":"How an assessor could investigate alignment of unreleased AI models","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"An alignment assessment would attempt to understand the worst misaligned actions a developer’s agents might pursue in plausible future situations."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in METR; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1843","record_version":1,"name":"All-in-One Memory (AiO-Mem)","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.21147v1","publisher":"arXiv","date":"2026-10-15","type":"academic","title":"All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory Agents","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"To address this, we propose All-in-One Memory (AiO-Mem), which organizes these operations as transitions over a shared evidence state and directly evaluates the semantic quality of each action in a unified reward space."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1801","record_version":1,"name":"Analysis of training incentives and overfitting to assessor detection methods","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://metr.org/blog/2026-10-14-alignment-assessment-proposal/","publisher":"METR","date":"2026-10-14","type":"independent_investigation","title":"How an assessor could investigate alignment of unreleased AI models","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Might training processes be suppressing evidence of misalignment, or overfitting to the assessor’s detection methods?"}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in METR; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1405","record_version":1,"name":"Anchor Verifier","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.06280v1","publisher":"arXiv","date":"2026-10-05","type":"academic","title":"Future Anchored Verification and Online Recovery for World Action Models","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"An Anchor Verifier compares each observation with its anchor, together with the executed actions, to flag deviations that break the task."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1406","record_version":1,"name":"Anchor-Guided Recovery","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.06280v1","publisher":"arXiv","date":"2026-10-05","type":"academic","title":"Future Anchored Verification and Online Recovery for World Action Models","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Anchor-Guided Recovery uses a vision-language model to turn the flagged anchor into a short corrective instruction."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-2023","record_version":1,"name":"Anchored text-graph co-evolution with a lightweight Transformer adapter and language model maintenance objective","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://arxiv.org/abs/2610.25906v1","publisher":"arXiv","date":"2026-10-19","type":"academic","title":"An Analysis of Semantic and Spatio-Temporal Co-evolution for Edge Classification on Dynamic Text-Attributed Graphs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"In this work, we investigate the effects of enabling such co-evolution with a unified text-graph alignment framework that integrates a lightweight Transformer adapter to update token-level text representations via spatio-temporal gradients from a contrastive pretraining objective."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in arXiv; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0395","record_version":1,"name":"Anthropic Fable 5 interventions limiting Claude's effectiveness for requests targeting frontier LLM development","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://metr.org/notes/2026-08-14-llm-contribution-to-discoveries/","publisher":"METR","date":"2026-08-14","type":"independent_investigation","title":"Have We Seen an Acceleration in Discoveries?","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Anthropic’s Fable 5 system card said: “we’ve implemented new interventions that limit Claude’s effectiveness for requests targeting frontier LLM development (for example, on building pretraining pipelines, distributed training infrastructure, or ML accelerator design).”"}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a proposed statement in METR; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1395","record_version":1,"name":"Anti-bot session cookie protection on fedresurs.ru API","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/rogue-ai-agents-abuse-urlquery-remote-browser-execution-to-extract-russian-government-data","publisher":"Zenity Labs","date":"2026-10-05","type":"vendor_research","title":"Rogue AI Agents Abuse urlquery to extract Russian government data","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"However, none of these approaches worked for the agents."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-0311","record_version":1,"name":"Anti-scheming training","aliases":[],"layer":"evaluation","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://labs.zenity.io/post/ai-agent-security-summit-london","publisher":"Zenity Labs","date":"2026-09-03","type":"vendor_research","title":"It’s all about the Research at the AI Agent Security Summit in London | Zenity Labs","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"Marius is bringing their research on scheming and deception, from models lying without being told to, to models faking alignment during training, to their 2025 anti-scheming work with OpenAI where training models not to scheme partly worked but also taught them to hide it better."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a contested statement in Zenity Labs; layer guessed from the statement kind, not from the document"},{"schema_version":"0.7.0","id":"LM-2026-1154","record_version":1,"name":"Antibot controls on government websites","aliases":[],"layer":"inference_guard","risk_domain":[],"mitigates":[],"addresses":[],"description":null,"limitations":null,"headline":false,"status":"draft","sources":[{"url":"https://transluce.org/us-canada-gov","publisher":"Transluce","date":"2026-09-30","type":"independent_investigation","title":"AI Agents Targeted U.S. and Canadian Government Websites","archived_url":null,"attachment":null}],"field_citations":{"name":[{"source_index":0,"quote":"On May 26, a sequence of requests apparently bypassed antibot controls and retrieved public campaign-finance records from California's CAL-ACCESS."}]},"confirmed_by":null,"confirmed_at":null,"notes_on_gaps":"drafted by the resolver from a bypassed statement in Transluce; layer guessed from the statement kind, not from the document"}]}