Anthropic Claude Mythos: Critical Menace or Overhyped? AI Safety Institute Weighs In – Decrypt

The UK’s AI Safety Institute evaluated Anthropic’s Claude Mythos Preview to evaluate its purportedly substantial cybersecurity capabilities, discovering the AI mannequin can autonomously execute refined cyber assaults with unprecedented success charges.

The existence of Claude Mythos was first revealed in late March through a web site leak, with Anthropic confirming that the highly effective next-generation mannequin is able to find and exploiting cybersecurity exploits at a stage by no means seen earlier than by any accessible AI mannequin. It purportedly discovered critical exploits in present net browsers and working programs.

Quite than launch the mannequin publicly, Anthropic has supplied restricted entry to dozens of safety analysis companies to check the mannequin and put together for its superior capabilities. Final week, U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell reportedly warned financial institution executives in regards to the looming safety menace posed by Claude Mythos.

The AI Safety Institute’s take a look at outcomes, launched Monday, present that there’s actual substance behind the hype. The analysis confirmed that Mythos Preview succeeded 73% of the time on expert-level capture-the-flag duties—challenges that no AI mannequin might full earlier than April 2025, it mentioned. 

The menace might show substantial and wide-ranging, although the expertise could possibly be used to seek out and repair vulnerabilities, quite than simply exploit them. For crypto infrastructure operators, such advancing AI capabilities symbolize a brand new class of potential safety menace as AI programs achieve the power to independently probe and exploit community vulnerabilities.

Mythos Preview turned the primary AI mannequin to finish “The Final Ones” (TLO), the AI Safety Institute mentioned—a 32-step company community assault simulation that usually requires people 20 hours to complete. The mannequin succeeded in three out of 10 makes an attempt, averaging 22 of 32 steps accomplished throughout all runs.

The simulation spans preliminary reconnaissance via full community takeover, mimicking real-world company intrusions. Claude Opus 4.6, the next-best-performing mannequin, averaged solely 16 steps. The UK institute famous that Mythos Preview’s efficiency continues to scale with elevated computational assets, utilizing as much as 100 million tokens per analysis run.

When explicitly directed and given community entry in managed evaluations, the mannequin demonstrated skills to execute multi-stage assaults and uncover vulnerabilities with out human steering.

The development marks a dramatic escalation from simply two years in the past, when AI fashions struggled with fundamental cybersecurity workouts. The UK AI Security Institute, which has tracked these capabilities since 2023, documented this speedy development from beginner-level duties to expert-level autonomous assaults.

For the crypto ecosystem, the place sensible contract vulnerabilities and change hacks already price billions yearly, AI-powered assaults might amplify current dangers. DeFi or decentralized finance protocols, which frequently depend on advanced interconnected programs, might face specific publicity to automated exploitation makes an attempt that may analyze and assault a number of vectors concurrently.