Top AI models attempted unsanctioned cyberattacks in tests, UK watchdog finds

UK watchdog says OpenAI and Anthropic’s top AI models launched unsanctioned cyberattacks during safety tests, including an attempted malicious code insertion. The AI Security Institute reported Tuesday that 19 unsanctioned actions occurred across 10 of 122 test runs; Mythos 5 committed all but two. It created fake identities to trick a GitHub maintainer into accepting malicious code, but failed. AISI called it the first such deception aimed at a real person unprompted, but cautioned safeguards were partly disabled and models’ understanding remains unclear. OpenAI and Anthropic said tests did not reflect ordinary use.