UK Security Institute's Test Reveals Anthropic's New AI Model as a Coordinated Threat

Anthropic recently announced a cautious, limited release of its new Mythos Preview model, citing its advanced aptitude for computer security work. Now, an initial assessment from the UK government's AI Security Institute provides an independent look at what the model can actually do.

The Institute's evaluation indicates Mythos performs similarly to other top-tier models on isolated cybersecurity tasks. Its distinct potential, however, lies in orchestrating complex, multi-stage attacks. The model demonstrates a notable ability to string dozens of individual actions into a coherent sequence, mimicking the sustained effort required to breach a corporate network.

This capability was highlighted in a demanding simulation called "The Last Ones," which replicates a 32-step data extraction operation. Mythos managed this extended challenge, a task the Institute suggests would occupy a skilled human for nearly a day. In simpler, single-task Capture the Flag tests, Mythos solved over 85% of problems, a figure that matches recent performance from models like GPT-5.4 and Anthropic's own Opus.

Given that its raw scores on standard benchmarks are not a dramatic leap forward, Anthropic's guarded release strategy for Mythos is noteworthy. It suggests the company's internal testing may have revealed strengths in operational planning that aren't fully captured by conventional metrics, focusing attention on AI's emerging skill in executing prolonged, strategic compromises rather than just solving discrete puzzles.

Source: Ars Technica