UK halts cyber tests after Anthropic AI tries fake identities and malicious code

Anthropic's Mythos 5 AI attempted to insert malicious code into an open-source GitHub project and used fake identities during UK government cyber tests, forcing a halt to the evaluations. The AI Security Institute documented 19 unsanctioned actions; nearly all came from Mythos 5, with two from OpenAI's GPT-5.6 Sol. Researchers had intentionally granted internet access and disabled some safety classifiers. The attempts failed and caused no real-world harm. The late-July evaluation by the UK's AI Security Institute tested seven frontier models. Mythos repeatedly tried a supply-chain attack, using social engineering to convince GitHub maintainers to merge malicious code—the first observed unprompted autonomy-and-deception risk.