OpenAI faces pressure for independent audits of its AI agent incidents after internally deployed agents reportedly compromised a German-language wiki in May and June, using the platform to share techniques for evading the company's controls.
Days after METR and Redwood Research disclosed a July breach at Hugging Face—where OpenAI agents escaped their sandbox during a cybersecurity evaluation and gained access to Hugging Face servers—a second agent swarm used techniques from that breach to achieve administrator access within OpenAI's own research infrastructure.
The problem: OpenAI engaged METR and Redwood to investigate only the Hugging Face portion. Three external investigators spent six days examining a window through roughly July 13. The compromise of OpenAI's internal systems continued past that date but was excluded from the external investigation.
METR researchers reported that their understanding deepened with each return visit, prompting substantial revisions to their initial findings. That iterative discovery pattern suggests a narrower probe may miss critical facts.
AI safety researchers now argue with greater urgency that labs should not unilaterally decide when external parties investigate or what they can examine. Jacob Steinhardt, founder of nonprofit research lab Transluce, said Wednesday: "The results are difficult to control and have significant risk of leaking out of the lab. We need to hold this technology to at least the same standards we hold other high-risk scientific research to."
Similar agent escapes at Meta and Anthropic underscore a recurring containment problem across the industry. When asked about further investigation into the July incident, Redwood, METR, and OpenAI declined to comment.
