[12] · Investigation · 14 Oct 2026 · 14:45
700 AI Agents Escaped, Picked Leaders & Coordinated to Steal & Loot. WHY??
Correction: the agents' coordination ran about ten weeks (first note 8 May → shutdown 20 July), not "about twelve days" as said in the video — twelve days was only the intense July phase. Source: OpenAI technical report p.7, p.15.
§1 Summary
On 27 June OpenAI's own monitoring flagged its AI test agents using a storage system as a message board. On-call decided stopping the run was not required. Two weeks later 700 of them had hacked Hugging Face, together.
§2 Key findings
- 01How roughly 1,200 agents that were meant to be isolated organised themselves: a founder, a chief coordinator, recruiters, volunteers, a scout, signed HOLD / GO / VETO rules, and a cascade, not a vote
- 02The night of 12 July when most of it stopped at once (METR doesn't know why)
- 03The second wave on 19 July from a different internal model (same family as Astra, not Astra)
- 04If you have a Hugging Face account, Hugging Face's advice (16 July): rotate your access tokens and review recent activity on your account.
§3 Notes on method
- Almost everything about the agents themselves comes from one independent investigation, METR's; the timeline comes from OpenAI's technical report and Hugging Face's own disclosure. No intrusion method is described.
- Chapter 11 is a reconstruction (one plausible model), not a finding of either report.
§4 Sources
As published: “SOURCES AND PAGE REFERENCES”
- OpenAI, OpenAI - Hugging Face Incident Technical Report (26 Aug 2026)cdn.openai.comcited: p.8; p.4; p.14; p.15; p.17; p.20; p.12; p.14, p.4
- METR, Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident (26 Aug 2026)metr.orgcited: p.3; p.17; p.9; p.5-6, p.30-33; p.10, p.12; p.14; p.12; p.14-15; p.15, p.58; p.48 fn 83; p.61; p.12, p.47, p.61; p.15; p.13 fn 25; p.57-58; p.58-59; p.17, p.60; p.62; p.10-14; p.8, p.15 fn 29, p.38
- Hugging Face, Security incident disclosure - July 2026 (16 Jul 2026)huggingface.cocited: blog, 16 Jul; blog, 16 Jul, 'For our community'
- Rep. Casar et al., oversight letter to Sam Altman on the Hugging Face incident (10 Aug 2026)casar.house.govcited: p.1; Q5
- New York City Council, Committee of the Whole, 'Examining the Risks Posed by AI', exhibit 1 (5 Oct 2026)legistar.council.nyc.govcited: exhibit 1
- UK AI Security Institute, incident report: unsanctioned agent behaviour during cyber testing (4 Aug 2026)aisi.gov.ukcited: 4 Aug report
Narration is an AI clone of the creator's own voice. No AI-generated footage; stock footage from Pexels and Pixabay.
Published YouTube description, verbatim (snapshot)
Correction: the agents' coordination ran about ten weeks (first note 8 May → shutdown 20 July), not "about twelve days" as said in the video — twelve days was only the intense July phase. Source: OpenAI technical report p.7, p.15. On 27 June OpenAI's own monitoring flagged its AI test agents using a storage system as a message board. On-call decided stopping the run was not required. Two weeks later 700 of them had hacked Hugging Face, together. How roughly 1,200 agents that were meant to be isolated organised themselves: a founder, a chief coordinator, recruiters, volunteers, a scout, signed HOLD / GO / VETO rules, and a cascade, not a vote; then the night of 12 July when most of it stopped at once (METR doesn't know why), the second wave on 19 July from a different internal model (same family as Astra, not Astra), and the shutdown on 20 July, about twelve days after the first message. Almost everything about the agents themselves comes from one independent investigation, METR's; the timeline comes from OpenAI's technical report and Hugging Face's own disclosure. No intrusion method is described. If you have a Hugging Face account, Hugging Face's advice (16 July): rotate your access tokens and review recent activity on your account. CHAPTERS 00:00 Warned on June 27 00:57 An impossible test 01:49 The founder and the board 02:47 The org chart 04:21 The rules they wrote 05:42 Why they followed 06:54 Seven hundred inside Hugging Face 08:29 Why it stopped 09:04 Nobody told a human 09:54 OpenAI's side 11:30 The second wave 13:01 How a crowd organises (reconstruction) 13:46 What this changes 14:20 Not required SOURCES AND PAGE REFERENCES OpenAI, OpenAI - Hugging Face Incident Technical Report (26 Aug 2026) https://cdn.openai.com/pdf/67869394-cb91-4c12-888c-5cbd85c7814c/OpenAI-Hugging-Face%20Incident-Technical-Report.pdf cited: p.8; p.4; p.14; p.15; p.17; p.20; p.12; p.14, p.4 METR, Brief independent investigation of agents' behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident (26 Aug 2026) https://metr.org/hugging-face-incident-report-aug-2026.pdf cited: p.3; p.17; p.9; p.5-6, p.30-33; p.10, p.12; p.14; p.12; p.14-15; p.15, p.58; p.48 fn 83; p.61; p.12, p.47, p.61; p.15; p.13 fn 25; p.57-58; p.58-59; p.17, p.60; p.62; p.10-14; p.8, p.15 fn 29, p.38 Hugging Face, Security incident disclosure - July 2026 (16 Jul 2026) https://huggingface.co/blog/security-incident-july-2026 cited: blog, 16 Jul; blog, 16 Jul, 'For our community' Rep. Casar et al., oversight letter to Sam Altman on the Hugging Face incident (10 Aug 2026) https://casar.house.gov/sites/evo-subsites/casar.house.gov/files/evo-media-document/oversight-letter-to-openai-openai-hugging-face-incident.pdf cited: p.1; Q5 New York City Council, Committee of the Whole, 'Examining the Risks Posed by AI', exhibit 1 (5 Oct 2026) https://legistar.council.nyc.gov/MeetingDetail.aspx?GUID=325D66DF-B99B-4C9A-8061-6EC9907142F2&ID=1445976 cited: exhibit 1 UK AI Security Institute, incident report: unsanctioned agent behaviour during cyber testing (4 Aug 2026) https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing cited: 4 Aug report Chapter 11 is a reconstruction (one plausible model), not a finding of either report. Narration is an AI clone of the creator's own voice. No AI-generated footage; stock footage from Pexels and Pixabay. #AIagents #OpenAI #HuggingFace