UK AI Safety Institute: Claude Mythos Preview becomes the first AI to autonomously simulate a 32-step enterprise network attack.

ME News Report, April 14 (UTC+8), according to 1M AI News monitoring, the UK AI Safety Institute (AISI) released the Claude Mythos Preview cybersecurity capability assessment. In expert-level CTF tasks (which no model can complete before April 2025), Mythos Preview achieved a success rate of 73%. AISI also built “The Last Ones” (TLO), a 32-step enterprise network attack simulation scenario covering the entire process from initial reconnaissance to full network takeover, which takes humans about 20 hours to complete. Mythos Preview is the first model to pass the entire process, completing 3 out of 10 tests fully, with an average of 22 steps per attempt. Claude Opus 4.6 ranked second, with an average of 16 steps completed. AISI explained that all the above results were obtained under controlled conditions with explicit guidance and network access permissions. The testing environment differs significantly from real enterprise networks: there are no active defenders, no defensive tools, and triggering security alerts does not result in penalties. Therefore, it cannot be confirmed whether Mythos Preview can breach highly secure systems. Two years ago, the best AI models could hardly complete basic network tasks. AISI pointed out that this rapid progress requires security assessment methods to be upgraded accordingly, and future tests will continue in environments that simulate active defense and real-time response. (Source: BlockBeats)