Safety news
New models, new publications, and regulation and policy, from the sources we watch.
New models 2
- 2026-09-29 · GPT-6.1 Sol · OpenAI · Preparedness Framework
- 2026-09-28 · Claude Sonnet 5.5 · Anthropic · Responsible Scaling Policy (RSP)
Publications 13
- 2026-09-30 · Unlearning Deceptive Behaviors in LLMs with Contrastive Forget Sets · arXiv
- 2026-09-30 · SteerProbe: Learning to Bypass Safety Steering in Vision-Language Models · arXiv
- 2026-09-30 · SceneJail: Exploiting Video Scenario Context to Jailbreak Multimodal LLMs · arXiv
- 2026-09-30 · Faithful Dual-constrained Erasure for Robust LLM Safety Alignment · arXiv
- 2026-09-30 · Chris Painter's testimony to the U.S. Senate on AI agent incidents · METR
- 2026-09-29 · pikit: A Composable Toolkit for Indirect Prompt Injection Research and Evaluation · arXiv
- 2026-09-29 · actr: aligning thoughts and responses for multilingual safety in reasoning llms · arXiv
- 2026-09-29 · Where Do LLMs Decide to Break the Rules? Mechanistic Localization of Prompt Injection Compliance · arXiv
- 2026-09-29 · The Unequal Influence of Bad Advice: Using Training Data Attribution to Modulate Emergent Misalignment · arXiv
- 2026-09-29 · Safer Content or Firmer Refusals? A Hybrid Perturbation Defense for Alignment under Harmful Fine-tuning · arXiv
- 2026-09-29 · Does the Unsafe Gradient Survive a Conversation? On the Fragility of Gradient-Based Jailbreak Detection in Multi-Turn Dialogue · arXiv
- 2026-09-29 · Controlled Decoding Attacks on Black-Box LLMs · arXiv
- 2026-09-29 · Can Vision-Language Models Stay Helpful When Facing Implicit Risks? Intent-Privilege OPSD for Efficient Safety-Helpfulness Alignment · arXiv
Regulation and policy in the news 2
- 2026-10-01 · AI agents could leave commercial losses between insurance policies · insurancebusinessmag.com
- 2026-09-29 · AI Quarterly | A Review of AI Law, Policy & Practice | 2026 Q3 | JD Supra · jdsupra.com
An item is listed because its source published it, with the date we hold; this is not a finding about it. Window: since 2026-09-25.
From Anthropic, Anthropic alignment, FAR.AI, Google AI, Google DeepMind, Google DeepMind safety, GreyNoise, Hugging Face, LessWrong, METR, OpenAI, Redwood Research, Schneier on Security, UK AI Security Institute, Zenity Labs, arXiv (safety). An item is listed only when its title names AI and also a safety, risk or policy matter, by a fixed keyword list; a title that only names the field ("AI safety") is not enough.
Agency publications themselves are on the Regulation tab; this lists press coverage. Selected by a fixed keyword rule on headlines from the news sources we watch, plus Cyberspace Administration of China. The match is mechanical and unreviewed.