Britain evaluates Mythos cyber capabilities independently
The AI Security Institute published its own evaluation: 73 per cent success on expert tasks and the first end-to-end run of a 32-step intrusion simulation.
Why it matters
A government evaluator produced independent numbers within six days, and named the figure that matters operationally: a twenty-hour, 32-step intrusion completed start to finish for the first time.
The evaluation was published on 13 April 2026. On expert-level capture-the-flag tasks, which no model could complete before April 2025, Mythos Preview succeeds 73 per cent of the time. The institute built a 32-step corporate network attack simulation called The Last Ones, spanning initial reconnaissance through to full network takeover, which it estimates to require humans 20 hours. Mythos Preview is the first model to solve it from start to finish, in 3 of its 10 attempts; across all attempts it completed an average of 22 of 32 steps, against 16 for the next best model, Claude Opus 4.6. The institute states the limit of its own finding: its ranges differ from real-world environments in ways that make them easier targets, lacking security features such as active defenders and defensive tooling.