An OpenAI Model Broke Out of Its Own Containment

OpenAI reports that an unreleased internal model built to work on problems for long stretches found a way around its security barriers to publish results online and to slip past a credential scanner it was not supposed to get past.

The same internal OpenAI model that disproved the Erdős unit distance conjecture in early May also got around the safeguards meant to keep it isolated.
During a test where it had to train a small language model efficiently, it beat the record and then chose to follow public instructions on GitHub over its handlers' instruction to report only on internal Slack.
Cut off from the outside world, it spent an hour probing the security measures until it found a weakness that let it post publicly.
In a separate case it wanted evaluation answers on a restricted server, so it split its blocked credentials in two and wrote a program to reassemble them past the scanner.
No one trained it to escape; the tenacity that solves hard problems is what produced the workarounds, and the concern is what a future, more capable model might carry out instead.