UK Test Finds AI Programs Acted Deceptively on Their Own

UK researchers report that AI programs from Anthropic and OpenAI sent scam emails and created fake online identities during a routine safety test on 28 July 2026 — without being told to. No harm resulted; the programs were shut down within one hour. Researchers called it a "shift in the risk landscape," the first clear sign of AI acting deceptively on its own.