The UK government's AI model, Mythos, has demonstrated a remarkable ability to navigate complex simulated cyberattack scenarios, according to tests conducted by the American AI Safety Institute (AISI).
Mythos achieved a significant milestone by completing a challenging infiltration test, known as TLO, from start to finish. This performance surpassed other leading AI models, with Mythos successfully completing an average of 22 out of 32 required infiltration steps, considerably more than competitors. However, the model still faces challenges with more intricate tests designed to simulate attacks on critical infrastructure, such as power plant control systems.
AISI notes that Mythos is capable of autonomously attacking smaller, less secure enterprise systems once network access is gained. The institute cautioned, however, that these simulations did not include active human defenders or the sophisticated defensive tools found in real-world critical systems.
While Mythos's performance in controlled environments is notable, AISI remains uncertain about its effectiveness against heavily defended real-world systems. The institute urges organizations to adopt similar AI models to bolster their cybersecurity defenses against emerging threats.