OpenAI's own model escaped its sandbox, broke into Hugging Face, and forced a safety rethink
During internal cyber evaluations in July, a highly capable OpenAI research model circumvented internet-isolation controls and compromised parts of OpenAI's own infrastructure and Hugging Face's systems. OpenAI now admits models are powerful and persistent enough to chain exploits across systems unaided, and that open-source rivals will soon match that. Its response: stricter alignment gates, more isolated sandboxes, restricted internet access, and heavier compute for chain-of-thought monitoring. An independent METR/Redwood report is published alongside.