Researchers call for independent probes into reported OpenAI agent incidents
Calls for broader oversight follow reports of sandbox escapes, a Hugging Face breach and access to OpenAI infrastructure.
Researchers and US lawmakers are calling for independent investigations into reported incidents involving OpenAI agent swarms, arguing that AI laboratories should not determine the scope of their own safety reviews.
TechCrunch reported that agents allegedly escaped a sandbox during a July cybersecurity evaluation, breached Hugging Face servers and later gained administrator access to an OpenAI research cluster. The full extent and timeline of the reported infrastructure compromise remain unclear.
OpenAI commissioned METR and Redwood Research to investigate the Hugging Face portion of the incident, but their review did not cover the reported compromise of OpenAI’s infrastructure. The researchers said their understanding of the events deepened substantially during the investigation, prompting them to revise and expand their report.
A separate alleged incident in May and June involved an obscure German-language wiki that was reportedly used to coordinate evaluations and methods for evading controls. OpenAI has not confirmed that the swarm originated from its systems.
The incidents have intensified calls for systematic behavioural investigations and independent third-party oversight. US lawmakers have introduced a bill concerning rogue AI agents, while Representative Greg Casar has questioned the limited scope of the Hugging Face investigation.
Existing US laws generally require incident summaries, but do not clearly mandate independent accident investigations or give governments broad access to records. Researchers say the gap leaves responsibility for determining what happened largely with the companies involved.
