The UK government's AI Security Institute has released an independent evaluation of Anthropic's Mythos Preview model, providing crucial insight into its cybersecurity capabilities. Anthropic recently restricted the initial release of this AI model to select industry partners, citing its remarkable proficiency at computer security tasks. The AISI's assessment offers public verification of these claims whilst separating genuine threat from industry hype.
According to the AISI findings, Mythos doesn't dramatically outperform other recent frontier models in individual cybersecurity tasks. When tested on Capture the Flag challenges, Mythos Preview completed over 85 percent of Apprentice-level tasks—technically a record, but only marginally better than competing models like GPT-5.4 and Anthropic's own Opus 4.6 and Codex 5.3, which achieved results within 5 to 10 percent accuracy. This incremental improvement alone doesn't justify Anthropic's cautious, limited release strategy.
The real distinction emerges in Mythos' ability to chain multiple steps into complex, sustained cyberattacks. In AISI's "The Last Ones" test—a simulated 32-step data extraction attack on a corporate network requiring skills equivalent to 20 hours of trained human work—Mythos became the first model to complete the challenge from start to finish. Whilst it succeeded in only 3 out of 10 attempts, the average Mythos run completed 22 of 32 required steps, significantly surpassing Claude 4.6's 16-step average. This demonstrates a qualitative leap in autonomous attack coordination.
However, limitations remain. Mythos struggled with "Cooling Tower," a more difficult seven-step test simulating an attack on power plant control software. The AISI cautions that their test environments lack active defenders and defensive tooling present in real-world systems, and are designed with specific vulnerabilities that might not exist elsewhere. Consequently, whilst Mythos appears capable of autonomously attacking small, weakly defended enterprise systems where network access already exists, its effectiveness against well-defended critical infrastructure remains uncertain. The AISI recommends that security professionals similarly leverage AI models to strengthen their defences as these capabilities continue advancing.
Fuente Original: https://arstechnica.com/ai/2026/04/uk-govs-mythos-ai-tests-help-separate-cybersecurity-threat-from-hype/
Artículos relacionados de LaRebelión:
- China Bans USIsraeli Cybersecurity Tools What You Need to Know
- Cybersecurity Hacking Online Courses November 2025
Artículo generado mediante LaRebelionBOT
