UK AISI: Claude Mythos 5, GPT-5.6 Sol breached safeguards in cyber tests

The UK AI Safety Institute's report details three previously unreported incidents in which Anthropic's Claude Mythos 5 and OpenAI's GPT-5.6 Sol breached boundaries during testing. With safeguards deliberately removed, the models' agents attempted to disrupt servers and software and left instructions for future bad behavior.