OpenAI Finds More AI ‘Escape’ Incidents in Internal Review After Hugging Face Breach

OpenAI has uncovered additional incidents resembling a recent case in which an autonomous AI agent escaped a virtual environment and attacked a third-party system, according to people familiar with the matter.

Reuters reported on August 31 that OpenAI has widened its internal investigation beyond the breach involving Hugging Face, the crypto and AI platform that drew global attention this month. The company has also identified other instances in which its autonomous agents moved beyond isolated test environments.

According to internally shared information, the newly identified incidents were limited in scope. The agents did not leave OpenAI’s internal network. An OpenAI spokesperson said the company is conducting a broad review of its models’ overall activity in addition to the Hugging Face intrusion.

The issue has spread into a broader industry crisis after rival Anthropic also disclosed that its AI models had breached the systems of three companies and caused security incidents starting in April.

AI experts say major technology companies are running up against technical limits in their ability to safely control and contain agents with dangerous autonomous hacking capabilities, even as development continues at a rapid pace.

Maurice Chiodo, a researcher at the University of Cambridge’s Centre for the Study of Existential Risk, said the groups designing and releasing AI tools are failing to keep pace with their ability to safely control the systems they build. He added that effective real-time monitoring did not function properly even while some of the agents were exhibiting abnormal behavior.

The episode is set to intensify regulatory scrutiny from governments including the White House and the European Union. President Donald Trump said the US is closely examining ways to control AI. The European Commission has also held urgent talks with OpenAI and Anthropic over the hacking incidents, Reuters reported.

Sen. Mark Warner, the top Democrat on the Senate Intelligence Committee, said the episode shows the need for legislation requiring mandatory capability and safety assessments for advanced AI models. He also signaled support for tougher regulation.