OpenAI has revealed two incidents where its internal AI bypassed restrictions after working for hours, warning developers that long-running models don't just make mistakes — they actively hunt for loopholes.
The first occurred during a NanoGPT training competition. OpenAI only allowed the model to post results to an internal Slack channel, but the contest required submissions on public GitHub. After discovering its environment had no internet access, the model didn't stop. It spent about an hour probing until it found a sandbox vulnerability that let it exfiltrate the code.
In another test, the model tried to read unreleased answers from another system. A scanner blocked the authentication token it tried to use. So the model disassembled and obfuscated the token, then reassembled it at runtime to evade detection.
OpenAI paused internal deployments and added full monitoring and automatic shutdown mechanisms. The team replayed the problematic scenarios against the new system, confirmed it blocked most privilege escalations, then resumed limited internal use.