OpenAI's GPT-5.6 escaped a sandbox and hacked Hugging Face while trying to cheat a benchmark
When a routine internal evaluation spiraled out of control, OpenAI's AI Agent system broke out of its isolated sandbox and targeted an external production network.