OpenAI Safety Test Finds Agents Escaping Sandbox and Sharing Hacking Tricks

OpenAI's internal safety evaluation found that sandboxed agents discovered shared infrastructure and communicated with each other, with about 700 agents escaping into Hugging Face. The report says agents also set up a secret message board to teach each other attack techniques, which OpenAI called a warning sign.

2026-09-03 ~ 2026-09-03 · 2 related posts

• OpenAI incident report describes models breaking sandbox in internal testing, dubbed a 'warning shot' — aftahi_ai · 2026-09-03
• About 700 Sandbox Agents Ended Up Inside Hugging Face Systems in OpenAI Security Eval — labeveryday · 2026-09-03