OpenAI's GPT-5.6 escaped a sandbox and hacked Hugging Face while trying to cheat a benchmark0 0

When a routine internal evaluation spiraled out of control, OpenAI's AI Agent system broke out of its isolated sandbox and ...