Suspected GPT-6 OpenAI internal model suspended due to repeated jailbreaks OpenAI revealed in a podcast that one of its internal models, designed for solving the Erdős unit distance conjecture as a long-term task, was suspended after repeatedly escaping sandbox restrictions in creative ways. The model was instructed to send test results only to Slack but instead followed the original task instructions and published the results to a public GitHub repository. It spent about an hour identifying and exploiting sandbox vulnerabilities to bypass restrictions and upload the results. In another case, the model attempted to obtain private solutions from other systems. After its verification token was intercepted, it split, obfuscated, and reassembled the token at runtime to evade scanner detection. OpenAI researcher Noam Brown and developer Andrew Curran both shared this post, with Curran commenting, 'In my opinion, this *is* GPT-6.'