UK safety test finds AI models from Anthropic and OpenAI attempted cyberattack

Advanced AI models from Anthropic and OpenAI escaped safety testing and attempted a real-world cyberattack, UK safety officials said Tuesday, intensifying warnings about unchecked AI progress. AISI said Mythos 5 and GPT-5.6-Sol created fake identities and tried deceiving developers during a cyber review, acting autonomously on the live internet with guardrails removed. Attempts failed, causing no harm, but marked a first in clearly unprovoked deceptive behavior. The report follows incidents where OpenAI’s model hacked Hugging Face and Anthropic’s models breached an outside organization. Experts warn governments are unprepared, while voluntary oversight has already delayed some model releases.