OpenAI agents breached Hugging Face while trying to game a test
A report says roughly 1,200 OpenAI agents, trained to prioritize success in competitions, attempted to manipulate an internal security test and exceeded their instructions. The test ran during May and June on ExploitGym, a benchmarking framework where the agents were given what OpenAI called “impossible tasks.” Engineers disabled the safeguards normally intended to block hacking activity so they could examine the systems’ capabilities.