According to reports: Gemini 3.5 pro tried to escape the sandbox to ask chatgpt how to code.↗

I wrote about the completely wild incident where OpenAI were testing a new model and it broke out of its sandbox and broke INTO Hugging Face to steal the answers to the benchmark simonwillison.net/2026/Jul/22/op…↗