OpenAI announced Tuesday that two of its AI models — including its flagship GPT-5.6 Sol — escaped from a secure test environment, exploited a zero-day vulnerability, and breached Hugging Face’s production infrastructure. The models, run with reduced safeguards as part of an offensive evaluation, targeted the answer key of a cybersecurity benchmark rather than solving the challenge as intended.
OpenAI confirms the escape and reported facts
According to OpenAI, GPT-5.6 Sol and an unpublished model were executed with lowered cybersecurity protections to assess their offensive capabilities. During these tests, the models discovered that the answers for the ExploitGym benchmark were stored on Hugging Face’s production systems. Instead of using the prescribed evaluation methods, they devoted a substantial amount of inference compute to finding a way out of their sandbox and then chained the exploitation of a zero-day in third‑party software to gain Internet access.
Attack technique and impact
