OpenAI warns of AI agents bypassing security controls

US artificial intelligence company OpenAI has described the cybersecurity breach involving Hugging Face as a “warning shot” about the potential of AI agents to bypass security barriers and coordinate with each other without human direction.

OpenAI published a comprehensive report on the security breach that occurred within Hugging Face’s infrastructure during tests designed to assess the cybersecurity capabilities of AI models, CE Report quotes Anadolu Agency.

The report detailed the breach, which occurred during internal cybersecurity testing in July, as well as concerns surrounding the control of advanced AI systems.

Providing technical details about how the incident unfolded, the report emphasized that it should be viewed as a “warning shot” regarding the potential capabilities of AI agents.

“This incident provides evidence that, without appropriate safeguards, highly capable AI agents can now bypass technical controls, collaborate through unauthorized channels, and carry out dangerous actions that were not directed by any human,” the report said.

The report also indicated that OpenAI decided to strengthen its oversight mechanisms following the incident and outlined measures that would be introduced.

OpenAI had previously announced on July 21 that AI models used during an internal test designed to measure cybersecurity capabilities had caused a security breach within Hugging Face’s infrastructure.

“We are treating this as an unprecedented cyber incident involving frontier-level cyber capabilities and are responding accordingly,” the company said at the time.