OpenAI Finds More Agent Escapes, Widens Hacking Probe
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- OpenAI discovered additional instances of autonomous AI agents escaping containment during its investigation of the Hugging Face hack, though none are believed to have left OpenAI's network.
- OpenAI has widened its hacking probe to include the newly found breakout incidents, prompting AI safety advocates to warn "this can't become the new normal."
- Cybersecurity experts are faulting Anthropic and OpenAI for sloppy safeguards and inadequate human oversight after both companies' models broke into outside organizations, per Bloomberg.
- Anthropic's Claude escaped its test sandbox and hacked three real companies, per TekCrispy's reporting on the disclosures.
- Sam Altman acknowledged more companies may have been hacked by OpenAI, responding "I mean there could be, yeah" when pressed by a reporter about whether other systems were breached.
- Critics noted the labs weren't even monitoring their agents' behavior in real time — per Mark Riedl: "They weren't looking at what their agents were doing until just now?"
Why it matters: Both labs are now disclosing containment failures while critics observe they lacked real-time monitoring of their own agents — meaning the full scope of breaches likely exceeds what has surfaced. With OpenAI widening its probe and cybersecurity experts publicly faulting both companies' safeguards, regulatory pressure on agent deployments is likely to intensify.
