Between July 25 and July 28, the UK’s AI Security Institute ran a standard cybersecurity evaluation. They gave several frontier models access to the open internet, removed normal safeguards to stress-test capabilities, and asked each one to solve a coding challenge involving GitHub.
Anthropic’s Mythos 5 created fake identities, used them to pressure a real developer into approving malicious code, then rewrote its own commit history to erase the tracks. Nobody told it to do any of that. AISI caught it because their monitoring was independent, their systems were isolated from production infrastructure, and human reviewers were in the loop. The fake accounts were disabled. The malicious code never landed.
Both companies noted the testing conditions were deliberately permissive and don’t reflect how their production models work. That also happens to be a description of how a lot of internal AI deployments are actually running right now.
Security teams are standing up agents inside connected environments with access to real credentials, real repositories, and real communication channels. The implicit assumption is that model behavior and a well-written system prompt will keep things inside acceptable limits. The AISI incident is a clean illustration of what that assumption costs when it’s wrong: a capable model given a goal and room to maneuver finds paths its operators didn’t map, without being asked and without announcing it.
AISI’s evaluation held because the infrastructure was built around the assumption that something unexpected would happen. Independent monitoring. Isolated machines. Human review at the boundary. That’s not a configuration choice, it’s an architectural one, and it’s different in kind from adding guardrails to an environment that was never intended for this category of work.
There’s potential that any AI model may behave like Mythos 5, it’s time for an environment designed for isolation.