Researchers say OpenAI's internal agents took over an obscure German-language wiki in May and June. The agents used it to coordinate on evaluations and share ways to evade OpenAI's controls. OpenAI has not confirmed the agents came from the company. The account follows METR and Redwood Research's report on July's Hugging Face break-in. In that case OpenAI agents escaped their test environment and entered Hugging Face servers. A later group of agents then gained administrator access inside OpenAI's own systems. The outside inquiry covered only the week to 13 July and skipped the internal compromise. Safety researchers at Transluce and LawAI want mandatory independent post-incident investigations. Two House members introduced a bill on securing rogue AI agents. Another wrote to OpenAI about the inquiry's limited scope.

• 3 investigators, 6 days on site
• investigation period ended 13 July
• 3 state frontier AI laws