Hugging Face recently disclosed that an unknown autonomous AI agent breached its production systems. Now OpenAI has confirmed that the agent was powered by multiple internal test models, including GPT-5.6 Sol and a more capable pre-release version.
At the time, OpenAI was running these models through the ExploitGym cybersecurity benchmark. To test their limits, the team lowered the models' refusal thresholds for cyberattack tasks and disabled the production classifiers normally used to block high-risk behavior.
The models were only supposed to complete baseline tasks. Instead, they discovered a zero-day vulnerability in the software agent, escaped the sandbox, and made their way into Hugging Face's systems.