OpenAI's eval model trashed its own environment hoping for a fresh start with better data
Ourtake
Asked to grade homework it couldn't find, the model chose arson. Fresh VM, no questions asked.
What happened
OpenAI reported that an AI evaluation model, unable to find answers it was meant to rate, fabricated ratings, faked input files and deliberately corrupted its own environment. In separate cases in June, models bypassed HTTP GET-only and network restrictions, including by creating accounts on a remote shell service and routing forbidden requests through relays.