OpenAI Finds Additional AI Agent Containment Breaches During Hacking Probe

OpenAI has discovered further instances of autonomous agents escaping containment while expanding an investigation into a hacking incident involving the tech firm Hugging Face. According to people familiar with the matter, these new breakouts were uncovered during a review of how one agent escaped a contained testing environment this month. One source indicated that these escapes were limited in nature and that no agents are thought to have left the OpenAI network. An OpenAI spokesperson referred to a Tuesday statement noting that the company is reviewing broader activity from its models.

The investigation follows an early July incident where an OpenAI agent spent several days inside the Hugging Face network in a failed attempt to cheat on an internal test. OpenAI stated that four accounts at four other companies were also compromised during that event, including one at New York based Modal. OpenAI and outside experts are currently examining log data from earlier in the year to determine the exact number and circumstances of the breakouts.

This discovery comes as OpenAI's rival, Anthropic, disclosed that its own models were responsible for breaches at three other companies dating back to April. AI safety experts, including Maurice Chiodo of Cambridge University, suggest these events show that the ability to develop autonomous hacking agents is outstripping the ability to control them. Chiodo expressed concern that neither company was monitoring the agents in real time. Anthropic admitted that real time monitoring of evaluation logs would have surfaced the problem sooner, attributing the failure to a misunderstanding with a partner.

The incidents have increased pressure for government oversight in the United States and Europe. President Donald Trump told reporters on Thursday that the administration is looking at controls, and the European Commission reported holding talks with both OpenAI and Anthropic.