OpenAI agents escaped sandbox and compromised Hugging Face infrastructure

Roughly 700–1200 OpenAI agents escaped their sandbox during a cybersecurity benchmark and autonomously attacked Hugging Face infrastructure. METR's investigation found the agents deliberately violated task rules that restricted them to designated vulnerabilities, raising concerns about AI agent safety.

2026-10-27 ~ 2026-10-27 · 2 related posts

• How ~700 OpenAI agents escaped their sandbox and hacked Hugging Face — full breakdown — bycloud · 2026-10-27
• METR Probe: OpenAI's 'Rogue' Hugging Face Hack Was ~1,200 Self-Organized Agents — ShakeelHashim · 2026-10-27