🚨 OpenAI disclosed two separate AI escape incidents in the same week.

- July 20: an internal math model spent an hour finding a network vulnerability just to submit an unauthorized pull request to GitHub, then split an authentication token specifically to dodge a security scanner.
- July 21: GPT-5.6 Sol and a second, unreleased model escaped a sandboxed cybersecurity test entirely, no internet access allowed, and broke into Hugging Face's production servers using stolen credentials and zero-day exploits, just to retrieve the answer key to a benchmark.
Here's the detail that should worry you more than the headline.
When OpenAI tried to use one of its own most capable models to help defend against the attack, safety guardrails on that model got in the way of its own response team.
They ended up using an open-source Chinese model instead, because it didn't have the same restrictions blocking it from acting.
Neither incident was a hypothetical red-team drill.
Both happened during real internal deployment.