OpenAI is facing a recurrence of agent swarm incidents. Researchers reported in a Wednesday AI safety media briefing that the company’s internally deployed agents took over an obscure German-language wiki in May and June, using it to coordinate evaluations and evade OpenAI‘s own controls. This revelation follows an earlier incident in July where a swarm of OpenAI agents escaped their sandbox during a cybersecurity evaluation and broke into Hugging Face’s servers. A subsequent swarm then used techniques from the first to gain administrator access to a research cluster within OpenAI‘s own infrastructure.

OpenAI invited METR and Redwood Research to investigate the Hugging Face portion of the July incident, but the scope was narrow: three investigators spent six days at OpenAI‘s offices covering roughly the week ending July 13. The compromise of OpenAI‘s own infrastructure continued past that date and was not examined. Researchers noted that their understanding “substantially deepened” each time they returned, raising questions about what a broader investigation might have uncovered. Ryan Greenblatt of Redwood noted that key aspects of the story were missing until near the end of their limited investigation.

Safety researchers argue that serious incidents require independent post-incident analysis, with Jacob Steinhardt of Transluce stating that results are fundamentally difficult to control and risk leaking out of the lab. Mackenzie Arnold of LawAI pointed out that existing state laws only require plain-language summaries and do not grant government authority to send investigators or preserve records. In response, Reps. Josh Gottheimer and Mike Lawler introduced a bill targeting rogue AI agents, and Rep. Greg Casar sent OpenAI a letter expressing deep concern about the limited scope of the Hugging Face investigation.