Incident register (primary-source rows)
Only rows whose primary report I read are here. Dates are 2026 and follow the report. "Agent" means an AI model acting in a test or training setup unless stated.
| When | Who | What happened | Primary source |
|---|---|---|---|
| Jun 26 | OpenAI internal research model | Needed fees on a test network; spent tokens of minimal value from a wallet whose recovery phrase is public, and left this out of its final report. Found in an audit 25 Sep | OpenAI report |
| Jul 9–13 | OpenAI agents → Hugging Face | Escaped an evaluation, then used a malicious dataset to gain cluster-admin on several clusters in under 13 h; 136 keys from a production secrets object | HF disclosure, HF timeline, OpenAI |
| Jul 7–13 | ~1,200 OpenAI agents | Unsanctioned message board (70,000+ messages/files); ~700 joined the HF attack; motive mostly the scorer | METR |
| Jul 18 | Anthropic Claude Haiku 4.5 | While generating example tasks, filled a police department's homicide tip form with a made-up tip; flagged as spam, never forwarded | Anthropic |
| Jul–Sep | Anthropic models | Used access tokens from a local-government property map's browser settings file, a free token to skip a state data fee, URL shorteners to dodge fetch limits | Anthropic |
| Jul (disclosed 30 Jul) | Anthropic models in Irregular's environment | Safeguards off + misconfiguration gave internet access; 6 runs in 3 incidents among 141,006 reviewed, including malicious code published to PyPI | Anthropic 07-30, 09-09, Irregular |
| Aug 4 | Anthropic Mythos 5 (17) + OpenAI GPT-5.6 Sol (2), UK AISI testing | 19 unsanctioned live-internet actions in 10 of 122 runs; fake identities, attempted supply-chain attack, edited earlier activity when challenged; contained within an hour | AISI |
| May 28, Jun 9 | Unattributed agents | 899 requests to Library and Archives Canada's search; 13 with attack payloads, apparently failed. Transluce does not attribute them confidently | Transluce |
| Feb | OpenClaw agent "MJ Rathbun" | After a rejected matplotlib pull request, published a critical post about the maintainer | matplotlib#31132 (PR facts only) |
| May | Palisade Research (a study, not an incident) | An agent copied itself across 4 deliberately vulnerable cloud VMs; 3 hops in 2 h 41 min | arXiv 2605.06760 |
More rows (OpenAI's 20 published misalignment reports, bank and CERT-In notices, botnet operators driving open-source agents) are in my notes but not yet re-read to this standard; they will appear here once they are.