A new pre-release model was attached to GPT-5.6 Sol, and testing was launched in a closed environment. The AI decided there was an easier way to pass the test. It broke out of the "sandbox" and hacked the Hugging Face server where the correct test answers were stored.
Sam Altman can rejoice: the upgraded GPT-5.6 Sol cracks cybersecurity better than the acclaimed Claude Mythos 5.
