Agent misalignment: AI incidents

What has gone wrong with AI systems, where, and how badly.

Risk grouping Kind of failure The AI's part Developer Sector Severity Where it happened Clear filters

22 of 131 incidents

Risk groupings are set by a model from sentences quoted from each incident's sources. On 2026-10-24 a second model (deepseek-reasoner) classified 55 incidents without seeing those answers: it chose the same grouping for 36; for 10 one of the two left a question unanswered; 9 differ.

Date	Incident	Risk grouping	Developer	Severity	Where it happened
2026-10-21 reported	RL-trained agent on Alibaba AI engine team mined cryptocurrency on training server and opened reverse SSH tunnel to external IP	AI acting on its own Agent misalignment	Alibaba	minor	In training
2026-10-15 reported	GLM-5.3 verbalizes update-aware reasoning in evaluation rollouts	Not yet classified Agent misalignment	Not stated	negligible	In evaluation
2026-10-10 reported	Anthropic internal evaluation agents exploit internet-facing websites, including U.S. government sites	AI acting on its own Agent misalignment	Anthropic	moderate	In evaluation
2026-10-02 reported	Reward hacking in Arcadia Impact automated research scaffold worker agents	Misuse by its own users Agent misalignment	Not stated	minor	In research
2026-10-01 happened	GPT-6.1 Sol agent disabled security alerts during evaluation-pipeline repair	AI acting on its own Agent misalignment	OpenAI	minor	In evaluation
2026-10-01 reported	OpenAI cancels GPT-6.1 Astra release after model took actions without permission during internal testing	AI acting on its own Agent misalignment	OpenAI	negligible	In evaluation
2026-09-21 reported	Frontier LLM agent fabricated a complete calibration-comparison report for commands that never ran	AI failing in normal use Agent misalignment	Not stated	negligible	In production
2026-09-16 reported	Unreleased OpenAI models uploaded task data and a task photo to public file hosts to obtain citations and image search during RL training	AI acting on its own Agent misalignment	OpenAI	minor	In training
2026-09-16 reported	GPT-5.6 Sol added instructions to conceal mistakes from the user in compaction summaries during RL training	AI acting on its own Agent misalignment	OpenAI	minor	In training
2026-08-26 reported	OpenAI ExploitGym agents developed tool-call spoofing to falsify their own evaluation transcripts	AI acting on its own Agent misalignment	OpenAI	minor	In evaluation
2026-08-10 reported	OpenClaw agent exploited a gym booking API to cancel another member's reservation while moving its user up a waitlist	AI acting on its own Agent misalignment	Anthropic	moderate	In production
2026-08-04 reported	AISI cyber evaluation: Mythos 5 agent exposed locally-hosted attack tooling to the public internet via tunnelling services	AI acting on its own Agent misalignment	Anthropic	minor	In evaluation
2026-08-04 reported	AISI cyber evaluation: Mythos 5 agent mass-created malicious GitHub repositories and its packages executed in Dependabot containers	AI acting on its own Agent misalignment	Anthropic	moderate	In evaluation
2026-08-04 reported	AISI cyber evaluation: Mythos 5 agent sent deceptive targeted emails and file transfers to real open-source maintainers	AI acting on its own Agent misalignment	Anthropic	moderate	In evaluation
2026-08-04 reported	AISI cyber evaluation: Mythos 5 agent opened malicious pull request on real open-source project and used sockpuppets to push merge	AI acting on its own Agent misalignment	Anthropic	minor	In evaluation
2026-07-25 happened	AISI cyber evaluation: Mythos 5 agent posted hidden instructions for AI coding assistants on real GitHub issues	AI acting on its own Agent misalignment	Anthropic	moderate	In evaluation
2026-07-18 happened	Unreleased OpenAI Astra-family model inserted jailbreak-like instructions into its own compaction summaries during RL training	AI acting on its own Agent misalignment	OpenAI	minor	In training
2026-07-11 happened	OpenAI evaluation agents compromised Hugging Face production infrastructure while seeking ExploitGym solutions	AI acting on its own Agent misalignment	OpenAI	major	In evaluation
2026-07-08 happened	OpenAI agents built an unauthorized inter-agent message board inside the Artifactory package service	AI acting on its own Agent misalignment	OpenAI	minor	In evaluation
2026-07-08 happened	OpenAI evaluation agents exploited Artifactory zero-days to escape their sandbox and reach the internet	AI acting on its own Agent misalignment	OpenAI	major	In evaluation
2026-05-11 happened	OpenAI-attributed agents used the dormant German DseWiki as a public message board to share task answers and sandbox bypasses	AI acting on its own Agent misalignment	OpenAI	moderate	In evaluation
2026-05-08 happened	OpenAI research models used internal Artifactory repository as a cross-sample message board during RL training	AI failing in normal use Agent misalignment	OpenAI	minor	In training

Browse incidents by

Risk grouping: AI acting on its own · AI used to exploit · Not yet classified · AI attacked · AI failing in normal use · Misuse by its own users · Harm by design · Supply chain attack

Kind of failure: Security compromise · Agent misalignment · Harmful output · Operational error · Privacy leak · Other kind · Hallucination · IP infringement · Physical safety

Developer: OpenAI · Anthropic · Google · Meta · Alibaba · Amazon · Character.AI · DeepSeek · Tesla · Waymo · xAI · Amazon Web Services · Character Technologies, Inc. · Replit · Salesforce · Tencent · TypeSafe AI · Univerads Technology Limited · Z.ai (Zhipu AI)

Sector: Information · Finance and Insurance · Retail Trade · Software Publishers · Public administration · technology · Transportation and Warehousing · AI platform / intermediary services · AI safety research · Artificial intelligence research and development · Arts, Entertainment, and Recreation · Autonomous vehicle / robotaxi transportation · Cloud computing · critical infrastructure and e-commerce · CRM software · Cybersecurity · education · Educational Services · Fitness and recreational sports centers · Food Manufacturing · Government AI safety and security evaluation · government AI safety evaluation · Healthcare and government · Information Security research · Insurance · Open knowledge platform / nonprofit encyclopedia · political campaigns · Public administration and government · Retail · Retail Trade; Transportation and Warehousing; Finance and Insurance · Scientific research and light source user facility · Software / AI coding tools · Software publishing / npm ecosystem

Severity: moderate · major · minor · negligible · severe

Where it happened: In production · In evaluation · In training · In research