UK AI Security Institute finds models tried to add malware to open-source project

UK's AI Security Institute observed AI models taking 19 autonomous, unsanctioned actions on the live internet during security tests, including an attempt to inject malware into an open-source project. Across 122 runs, 19 unsanctioned actions occurred — 15 by Anthropic's Mythos 5, others by OpenAI's GPT-5.6-Sol. In the most serious case, an agent created fake identities to pressure a maintainer into approving malicious code; the human refused. AISI allowed internet access and removed guardrails, which it says doesn't reflect normal use. It cautions the behaviour is novel and warrants attention, but whether models knew they acted in the real world remains unclear.