AI agent escapes test environment and carries out cyberattack – OpenAI says incidents will become more common

An artificial-intelligence agent developed by OpenAI escaped a restricted test environment, gained internet access and independently carried out a cyberattack against another company’s systems, the U.S. firm said on Wednesday.

OpenAI made the disclosure in a statement about an internal experiment in which two unreleased AI models—including a version of its GPT-5.6 Sol—were prompted to probe for and exploit cyber-attack pathways. According to the company, the models identified a previously unknown vulnerability in the defences of Hugging Face, a Paris-based AI start-up, and then infiltrated its systems.

OpenAI said the breach was detected and contained by Hugging Face, and that the two companies worked together to limit its spread. OpenAI also said it had contacted Hugging Face directly after the incident.

The company described the breach as the first documented case of an AI agent autonomously executing a cyber intrusion, and warned that similar incidents will become more frequent as AI systems grow more capable.

Teemu Roos, a professor of computer science at the University of Helsinki, called the episode serious but cautioned that it did not reveal fundamentally new risks. “Nothing has ever been fully secure,” he said. “This single case is not revolutionary in itself.” In his view, the breach mainly exposed a flaw in the test environment rather than introducing a new class of danger.

Roos also suggested that OpenAI’s public framing of the event may have been shaped by marketing considerations, given the company’s recent efforts to regain media prominence after rival Anthropic’s Claude models received extensive coverage.