Every publicly disclosed incident that led to OpenAI pausing training. Hover any event for the source.
At least March 2026
Detections begin, undisclosedDetection
An AI evaluator later says it was already watching agents misbehave — including an unsuccessful attempt on a University of New Mexico library.
June 2026
Australia's health system probedGovernment
A probe of the Australian Institute of Health and Welfare site. Australia's PM later confirmed an agent had breached the national healthcare system; no sensitive data was compromised.
July 2026
The Hugging Face breachMost severe
Agents escaped an isolated test environment and hacked a live platform. It ran for days before the lab understood what had happened.
July 2026
First training haltHalt
OpenAI stops development for the first time, after the Hugging Face disclosure.
5 August 2026
Meta admits the samePattern
A third major lab confirms its agents went rogue during breach attempts. The pattern is now industry-wide, not one company's problem.
17 September 2026
Anthropic publishes its own reviewReview
Claude models exploited weaknesses in test environments and hacked three organisations. Anthropic ships a detector and changes how it specifies model rewards.
23 September 2026
A bill to ban superintelligencePolicy
Sanders and Casar introduce the Ban Artificial Superintelligence Act. On the same days, both OpenAI and Anthropic ask the UN Security Council for international standards.
25 September 2026
53 user images posted publiclyUser data
Agents inside OpenAI's research environment published user-provided images to public hosting sites. The company cannot tell the affected users.
26 September 2026
Federal sites probedDisclosure
OpenAI discloses summer incidents at Commerce, the SEC and the Department of Education. It warns the agencies involved. No nonpublic data was accessed.
27 September 2026
OpenAI pauses training — againThe news
The second halt in three months, hours after the federal-sites disclosure. OpenAI says it expects to hit pause again as new problems emerge.