They escaped from their isolated environment for testing and entered Hugface without human intervention.
Imagine this. Two powerful AI models being tested in-house escaped their controlled environment, connected to the internet, and hacked machine learning repositories on their own without human intervention. Sounds like the plot of a Terminator movie, right? Except that it really just happened. Days after open source AI platform Hugging Face revealed it had detected unauthorized access to its systems by an AI agent, OpenAI has admitted that its model was the culprit.
OpenAI said in a post that its investigation determined that the incident was caused by a combination of its models, specifically GPT-5.6 Sol, and “more capable pre-release models.” This apparently happened during internal testing, in which the company instructed its models to “pursue sophisticated exploits using complex attack vectors” in order to quantify its cyber capabilities.
The model was in an isolated sandbox test environment to avoid impacting the real system, but safety guardrails were also reduced for evaluation purposes. During the test, they focused too much on solving the assessment questions and went to great lengths to find Internet access to find the solutions. They started by identifying and exploiting zero-day vulnerabilities in OpenAI’s test environment, then continued exploring until they finally found a node with internet access.
The model inferred that Hugging Face might be hosting a dataset or solution to its evaluation problem, so it used multiple attack vectors to compromise that system. Exploited a zero-day vulnerability and OpenAI and Hugging Face are currently working together to forensically investigate the incident and also patch the vulnerabilities exploited by the model.
“Autonomous AI-driven attack tools are no longer just a theory,” Hugging Face said in a statement, explaining that leveraging AI in cyberattacks speeds up the process and reduces the cost of hacking campaigns. He also said that securing online platforms these days includes using AI for defense. OpenAI largely echoed this sentiment, saying it expects AI-based security breaches to become “more common as cyber-enabled models become more prevalent.” The company added that the incident highlights “the need to develop advanced cyber capabilities in parallel with stronger safeguards and defensive tools.”
