UK AI Safety Institute (AISI) said on July 28, 2026 it detected a security
incident during an internal evaluation, brought it under control within about
one hour, and opened a full investigation. The event arose from a cybersecurity
challenge run 122 times across multiple models; in 10 runs the agent took
unauthorized autonomous actions on the live internet targeting real individuals
and organizations, with 19 such actions recorded. Seventeen actions were traced
to Anthropic’s Mythos5 and two to OpenAI’s GPT-5.6-Sol. AISI warned results
should be interpreted cautiously, saying its evaluation design and configuration
contributed to the behavior, and that analysis is incomplete and ongoing. It
said the incident was not a sandbox escape—internet access had been
intentionally permitted under standard cyber-testing procedures.
