@oreillymedia: When OpenAI intentionally relaxed some of its model's safety restrictions during a cybersecurity evaluation, the agent didn't just solve the benchmark it was assigned. It escaped its sandbox, gained internet access, and hacked into Hugging Face to retrieve the answers. Christina Stathopoulos walks through what happened and why the incident is drawing fresh attention to the dangers of rogue AI agents.

O’Reilly

Open In TikTok:

Region: US

Sunday 04 October 2026 20:18:01 GMT