Where reports come from

OpenAI Misalignment Reports and Notices

OpenAI's own incident write-ups: each report has an incident date, a narrative with transcripts, and the lab's response. 15 reports and 3 notices as of 2026-10-12. This is the model for what a lab-run disclosure page looks like.

Visit ↗ · Posts as @OpenAI

Anthropic Alignment Science Blog

Research posts, including case studies of agentic misalignment and risk reports. Mostly simulations and audits, not live incidents, so each post is labelled accordingly.

Visit ↗ · Posts as @AnthropicAI

Anthropic News (incident posts)

Where Anthropic publishes incident disclosures, for example 'An alignment assessment of recent cybersecurity incidents' (2026-09-09) and 'Investigating incidents in cybersecurity evals' (2026-07-30).

Visit ↗ · Posts as @AnthropicAI

UK AI Security Institute Blog

Independent government testers. Publishes incident reports from its own evaluations, for example the 2026-08-04 incident report on unsanctioned agent behaviour during cyber testing.

Visit ↗ · Posts as @AISecurityInst

METR Blog and Agent Incidents Page

Nonprofit evaluator. Maintains a chart of 44 documented AI agent incidents (https://metr.org/agent-incidents/) and published the Frontier Risk Report (2026-05-19). Also runs independent investigations of lab incidents.

Visit ↗ · Posts as @METR_Evals · Bluesky: metr.org; Substack: metr.substack.com

Apollo Research Blog

Tests frontier models for scheming and deception. Most posts are evaluation findings rather than live incidents.

Visit ↗ · Posts as @ApolloResearch

Palisade Research Blog

Demonstrations of models resisting shutdown or gaming tasks, plus a podcast that covers AI hacking incidents.

Visit ↗ · Posts as @PalisadeAI · YouTube: @palisaderesearch; Threads/Instagram: palisade.research

OpenAI Alignment Research Blog (RSS)

The general research blog of the same team. The RSS feed does not list the misalignment reports, only broader research posts, so it is watched separately.

Visit ↗ · Posts as @OpenAI

News search (Bing News RSS)

Used to find press coverage of incidents for the 'Reported elsewhere' page. The real article address is pulled out of Bing's redirect link, so every item links straight to the outlet. Every item is classified by the model first, and press items are always labelled secondary.

Visit ↗

METR on Bluesky

Public Bluesky timeline, readable without an account. Used to catch announcements before the blog post.

Visit ↗ · Posts as @METR_Evals

Google DeepMind (no incident page found)

No dedicated incident or misalignment-report page found. Google confirmed the May 2026 Gemini test incident only after press questions and said it was not misalignment. Covered through press sources until Google publishes its own.

Visit ↗ · Posts as @GoogleDeepMind

Meta (no incident page found)

Meta told Reuters about the August 2026 Muse Spark test incident and said it would publish more once it had all the facts. No Meta-published report found yet. Check again when one appears.

Visit ↗ · Posts as @AIatMeta

Irregular (evaluation firm)

Third-party security-testing firm named in several 2026 incidents. Worth watching for its white paper.

Visit ↗ · Ran the cyber tests behind the Anthropic, Meta and Google incidents. Said it is writing a white paper on running these evaluations safely. Site and handle not yet checked by hand.

Palisade Research @PalisadeAI

Confirmed from the Palisade site footer.

Apollo Research @ApolloResearch

Confirmed from the Apollo site footer.

METR @METR_Evals

Confirmed from the METR site footer. Also on Bluesky as metr.org.

UK AI Security Institute @AISecurityInst

Confirmed from the AISI site footer.

Anthropic @AnthropicAI

Official lab account. Handle not re-checked from a lab page this session.

OpenAI @OpenAI

Official lab account. Handle not re-checked from a lab page this session.