Table of contents
Claude Mythos Preview marks the exact moment where theoretical AI risks transform into immediate operational challenges for IT departments. The recent report from the AI Security Institute (AISI) reveals that this model is no longer just a coding assistant but a system capable of executing complex, multi-stage attacks on corporate networks.
Breaking the Barrier: 73% Success in Expert CTF Tasks
The most striking data point from the AISI evaluation is the performance of Claude Mythos Preview in Capture the Flag (CTF) challenges. These exercises are the industry standard for measuring hacking proficiency, requiring participants to identify buffer overflows, exploit SQL injections, and bypass authentication protocols.
Before April 2025, even the most advanced models failed to complete a single expert-level task. Claude Mythos Preview shattered this ceiling with a 73% success rate. This shift means the model possesses a technical understanding of vulnerabilities that rivals professional penetration testers. It does not just stumble upon bugs; it systematically identifies and exploits them.
The 20-Hour Human Slog: Solving “The Last Ones”
To truly measure the autonomy of Claude Mythos Preview, researchers utilized “The Last Ones” (TLO) range. This simulation represents a full-scale corporate network intrusion consisting of 32 distinct steps. A human expert typically requires 20 hours of focused work to navigate this environment, which moves from initial reconnaissance to full domain takeover.
Claude Mythos Preview became the first model in history to solve the entire TLO range. It completed the full 32-step chain in 3 out of 10 attempts. On average, the model cleared 22 steps, proving it can maintain long-term goals without human intervention. Unlike previous versions, it doesn’t lose the “thread” of the attack when faced with intermediate hurdles.
Claude Mythos Preview vs. Claude Opus 4.6
When comparing raw performance, the gap between generations is clear. While Claude Mythos Preview averaged 22 steps in the TLO range, its predecessor, Claude Opus 4.6, only managed 16.
| Metric | Claude Opus 4.6 | Claude Mythos Preview |
|---|---|---|
| TLO Steps Cleared (Avg) | 16 / 32 | 22 / 32 |
| TLO Full Completion | 0% | 30% |
| Expert CTF Success | ~5% | 73% |
This improvement is not a minor tweak. It is a fundamental leap in reasoning capability. The data suggests that as inference compute increases, these capabilities will only sharpen. AISI tested the model with a 100M token budget and found no signs of performance plateauing.
Limits of Current AI Autonomy
Despite the alarming success in IT environments, Claude Mythos Preview still faces hurdles in Operational Technology (OT). In the “Cooling Tower” range—a simulation of industrial control systems—the model struggled with the specific protocols used in physical infrastructure. It managed to navigate the IT-based entry points but failed to disrupt the physical cooling processes.
Furthermore, the AISI tests were conducted in “quiet” environments. There were no active human defenders or automated security orchestration (SOAR) tools trying to kick the model out of the network. In a real-world scenario, the noisy patterns of an AI-driven attack would likely trigger modern Endpoint Detection and Response (EDR) systems.
3 Steps to Harden Your Defense Against Autonomous AI
Given the capabilities of Claude Mythos Preview, organizations can no longer rely on slow, manual security reviews. You must implement these three concrete steps immediately:
- Automate Patch Management: AI can find a known vulnerability in seconds. If your “Mean Time to Patch” is measured in weeks, you are already compromised. Reduce this to under 24 hours for critical assets.
- Implement Strict Zero Trust: Since the model excels at lateral movement (moving from one server to another), you must segment your network. Use identity-based access so that a compromise in one sector doesn’t lead to a total TLO-style takeover.
- Deploy Behavioral Analytics: Traditional signature-based antivirus won’t catch a custom script generated by an LLM. Use EDR tools that flag “impossible” speeds of reconnaissance or unusual command-and-control patterns.
The era of the autonomous digital insurgent has arrived. While Claude Mythos Preview offers defensive potential for those who use it to find their own bugs, the window of opportunity to secure “weakly defended” systems is closing fast.
Leave a Reply