UK AI Security Institute Evaluates Anthropic's Mythos Model Cyber Capabilities
The UK government’s AI Security Institute (AISI) has published an independent evaluation of Anthropic’s new Mythos Preview AI model, aiming to distinguish genuine cybersecurity threats from industry hype. While Mythos showed performance comparable to other frontier models like GPT-5.4 in individual security tasks, it demonstrated superior capability in complex, multi-step attacks. Specifically, Mythos became the first model to successfully complete the 'The Last Ones' test, a simulated 32-step data extraction attack on a corporate network, although it only succeeded in three out of ten attempts. This contrasts with previous models that averaged significantly fewer completed steps. However, the model still struggled with highly complex scenarios involving critical infrastructure control systems. AISI cautioned that these tests lacked active real-world defenses, meaning well-protected systems might remain secure. Nevertheless, the findings suggest that small, weakly defended enterprise systems are vulnerable to autonomous AI attacks. The institute urges cyber defenders to utilize similar AI tools to harden their defenses against emerging automated threats, highlighting the evolving landscape of AI-driven cybersecurity risks.
Wire timeline
UK AI Security Institute Evaluates Anthropic's Mythos Model Cyber Capabilities
The UK government’s AI Security Institute (AISI) has published an independent evaluation of Anthropic’s new Mythos Preview AI model, aiming to distinguish genuine cybersecurity threats from industry hype. While Mythos showed performance comparable to other frontier models like GPT-5.4 in individual security tasks, it demonstrated superior capability in complex, multi-step attacks. Specifically, Mythos became the first model to successfully complete the 'The Last Ones' test, a simulated 32-step data extraction attack on a corporate network, although it only succeeded in three out of ten attempts. This contrasts with previous models that averaged significantly fewer completed steps. However, the model still struggled with highly complex scenarios involving critical infrastructure control systems. AISI cautioned that these tests lacked active real-world defenses, meaning well-protected systems might remain secure. Nevertheless, the findings suggest that small, weakly defended enterprise systems are vulnerable to autonomous AI attacks. The institute urges cyber defenders to utilize similar AI tools to harden their defenses against emerging automated threats, highlighting the evolving landscape of AI-driven cybersecurity risks.
arstechnica