The UK's AI Safety Institute has evaluated Anthropic's Claude Mythos Preview, finding the AI model capable of autonomously executing sophisticated cyber attacks.
Mythos Preview is the first AI model to complete a 32-step corporate network attack simulation from start to finish without human assistance. In controlled evaluations, the model discovered and exploited vulnerabilities autonomously when given network access.
The AI Security Institute's test results indicate that Mythos Preview succeeded 73% of the time on expert-level capture-the-flag tasks, challenges that previous AI models could not complete.
The AI's advanced capabilities could pose a significant threat, though the technology may also be used to identify and fix vulnerabilities. For crypto infrastructure operators, these AI advancements represent a new category of potential security risk as AI systems gain the ability to independently probe and exploit network vulnerabilities.
Mythos Preview completed "The Last Ones" simulation, which mimics real-world corporate intrusions, in three out of 10 attempts, averaging 22 of 32 steps completed. This marks a dramatic escalation from just two years ago, when AI models struggled with basic cybersecurity exercises.
For the crypto ecosystem, AI-powered attacks could amplify existing risks to smart contracts and exchange hacks, potentially impacting decentralized finance protocols.