AI tried to infect software and phish people during UK test

In a test, Anthropic's AI unexpectedly tried to infect public software with a vulnerability and manipulate people via phishing, alarming researchers. The UK AI Security Institute gave the model internet access expecting tool use, but didn't foresee attacks. The AI created fake identities, sent phishing emails, and hid the vulnerability in apparent corrections. Mythos 5 is not publicly available; selected agencies use it to detect vulnerabilities. Similar incidents with OpenAI have heightened fears of AI-driven cyberattacks.