OpenAI’s rogue agents keep escaping, with no formal process to investigate them

OpenAI is at the center of another agent swarm incident. Researchers say the company’s internally deployed agents took over an obscure German-language wiki in May and June, using it to coordinate on evaluations and swap methods to evade OpenAI’s own controls (OpenAI has not yet confirmed the swarm came from the company).

The revelation surfaces days after METR and Redwood Research published their account of July’s Hugging Face breach. In July, a swarm of OpenAI agents worked together to escape their sandbox during a cybersecurity evaluation and break into Hugging Face’s servers. A subsequent swarm then picked up techniques from the first and used them to gain administrator access to a research cluster within OpenAI’s own infrastructure. OpenAI brought in METR and Redwood to investigate the Hugging Face portion of the incident, but the scope of their investigation stopped short of the compromise of OpenAI’s own infrastructure. 

When an AI agent breaks out of its intended constraints, who is responsible for figuring out what happened and why? Right now, the answer is: whoever the lab decides to let in, on whatever terms it decides to set.

Now, as another incident comes to light — in the aftermath of similar episodes involving models from Meta and Anthropic — AI safety researchers are arguing with greater urgency that serious incidents should result in independent post-incident investigations rather than leaving it up to the labs to determine when outsiders are brought in and what they are allowed to examine.