Anthropic's Claude Mythos Preview has crossed a significant threshold in AI capabilities: it can now execute multi-stage cyberattacks autonomously on vulnerable networks, completing complex attack simulations that would take human security professionals days to finish. The UK's AI Security Institute (AISI) conducted rigorous evaluations and found that Mythos Preview represents a dramatic leap forward in AI cyber capabilities, succeeding at expert-level hacking challenges that no previous model could complete .

What Exactly Can Claude Mythos Do in a Cyberattack Scenario?

The AISI tested Mythos Preview on increasingly difficult cybersecurity challenges, starting with capture-the-flag (CTF) exercises where AI models must identify and exploit system vulnerabilities to retrieve hidden "flags." On expert-level CTF tasks, which no model could solve before April 2025, Mythos Preview succeeded 73% of the time . But the real breakthrough came with more complex, real-world-style attacks.

Researchers built a simulation called "The Last Ones" (TLO), a 32-step corporate network attack that spans from initial reconnaissance through complete network takeover. The AISI estimates this attack would require human security professionals approximately 20 hours to complete. Mythos Preview became the first AI model to solve the entire simulation from start to finish, succeeding in 3 out of 10 attempts. Across all attempts, the model completed an average of 22 out of 32 steps. The next best-performing model, Claude Opus 4.6, completed an average of only 16 steps .

This represents a qualitative shift in AI capabilities. Previous evaluations tested isolated skills, but real-world cyberattacks require chaining dozens of steps together across multiple network segments, sustained operations that demand hours or days of human expertise. Mythos Preview can now do this autonomously when given explicit direction and network access.

How to Protect Your Organization Against AI-Assisted Cyberattacks?

• Apply Security Updates Immediately: The AISI emphasized that regular application of security patches is critical, as Mythos Preview exploited known vulnerabilities in test environments. Organizations that delay patching are significantly more vulnerable to AI-assisted attacks.
• Implement Robust Access Controls: Limiting who can access which systems and data reduces the attack surface available to autonomous AI models. The research showed that models struggle when they cannot freely move between network segments.
• Enable Comprehensive Logging and Monitoring: Active monitoring, endpoint detection, and real-time incident response systems were absent from the test environments, making them easier targets. Organizations should deploy these defensive tools to detect suspicious AI-driven behavior patterns.
• Harden Network Configurations: Security configuration best practices, including network segmentation and firewall rules, make systems significantly harder for AI models to compromise, even when they have initial access.

Why This Matters More Than Previous AI Breakthroughs

The AISI has tracked AI cyber capabilities since 2023, progressively building harder evaluations to keep pace with AI progress. Two years ago, the best available models could barely complete beginner-level cyber tasks. The speed of improvement has been striking . What makes Mythos Preview's achievement significant is not just the performance jump, but the type of task it can now complete: sustained, multi-step operations that require planning, adaptation, and autonomous decision-making across complex systems.

However, the researchers included important caveats. The test environments lacked security features commonly found in real-world systems, such as active defenders and defensive tooling. There were no penalties for actions that would trigger security alerts. This means Mythos Preview's success on poorly defended systems does not guarantee it could attack well-defended enterprise networks with modern security infrastructure .

"Mythos Preview's success on one cyber range indicates that it is at least capable of autonomously attacking small, weakly defended and vulnerable enterprise systems where access to a network has been gained," the AISI stated in its evaluation.

AI Security Institute, UK Government

The researchers also noted that Mythos Preview's performance continues to improve with more computational resources. The cyber ranges were run with a 100-million-token budget, and the model's performance scaled up to that limit, suggesting further improvements would occur with additional computing power .

What's Next for AI Cybersecurity Evaluations?

The AISI plans to evolve its testing methodology to keep pace with AI capabilities. Future evaluations will include hardened and defended environments with active monitoring, endpoint detection, and real-time incident response systems. Researchers will also track how AI-enabled vulnerability discovery and penetration testing campaigns perform on real-world systems .

The dual-use nature of these capabilities is significant. While Mythos Preview's abilities pose security challenges, the same AI capabilities could help defenders identify vulnerabilities and strengthen their systems. The AISI recently released guidance with the UK's National Cyber Security Centre (NCSC) on how cyber defenders can both harness and prepare for frontier AI .

For organizations, the message is clear: investment in cyber defense is now urgent. The capabilities demonstrated by Mythos Preview suggest that future frontier models will be even more capable, making proactive security measures essential rather than optional.