Anthropic Claude Mythos: Serious Threat or Overhyped? AI Security Institute Weighs In

The UK's AI Security Institute evaluated Anthropic's Claude Mythos Preview to measure its purportedly important cybersecurity capabilities, uncovering the AI exemplary tin autonomously execute blase cyber attacks with unprecedented occurrence rates.

The beingness of Claude Mythos was archetypal revealed successful precocious March via a website leak, with Anthropic confirming that the almighty next-generation exemplary is susceptible of uncovering and exploiting cybersecurity exploits astatine a level ne'er seen earlier by immoderate disposable AI model. It purportedly recovered superior exploits successful existent web browsers and operating systems.

Rather than merchandise the exemplary publicly, Anthropic has offered constricted access to dozens of information probe firms to trial the exemplary and hole for its precocious capabilities. Last week, U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell reportedly warned slope executives astir the looming information menace posed by Claude Mythos.

The AI Security Institute’s trial results, released Monday, amusement that there’s existent substance down the hype. The valuation showed that Mythos Preview succeeded 73% of the clip connected expert-level capture-the-flag tasks—challenges that nary AI exemplary could implicit earlier April 2025, it said. 

The menace could beryllium important and wide-ranging, though the exertion could beryllium utilized to find and hole vulnerabilities, alternatively than conscionable exploit them. For crypto infrastructure operators, specified advancing AI capabilities correspond a caller class of imaginable information menace arsenic AI systems summation the quality to independently probe and exploit web vulnerabilities.

Mythos Preview became the archetypal AI exemplary to implicit "The Last Ones" (TLO), the AI Security Institute said—a 32-step firm web onslaught simulation that typically requires humans 20 hours to finish. The exemplary succeeded successful 3 retired of 10 attempts, averaging 22 of 32 steps completed crossed each runs.

The simulation spans archetypal reconnaissance done afloat web takeover, mimicking real-world firm intrusions. Claude Opus 4.6, the next-best-performing model, averaged lone 16 steps. The UK institute noted that Mythos Preview's show continues to standard with accrued computational resources, utilizing up to 100 cardinal tokens per valuation run.

When explicitly directed and fixed web entree successful controlled evaluations, the exemplary demonstrated abilities to execute multi-stage attacks and observe vulnerabilities without quality guidance.

The advancement marks a melodramatic escalation from conscionable 2 years ago, erstwhile AI models struggled with basal cybersecurity exercises. The UK AI Safety Institute, which has tracked these capabilities since 2023, documented this accelerated progression from beginner-level tasks to expert-level autonomous attacks.

For the crypto ecosystem, wherever smart contract vulnerabilities and exchange hacks already outgo billions annually, AI-powered attacks could amplify existing risks. DeFi oregon decentralized concern protocols, which often trust connected analyzable interconnected systems, whitethorn look peculiar vulnerability to automated exploitation attempts that tin analyse and onslaught aggregate vectors simultaneously.