United Kingdom - Ekhbary News Agency
The UK government's AI Security Institute (AISI) has published a crucial evaluation of Anthropic's Mythos AI model, shedding light on its advanced cybersecurity capabilities. This assessment follows Anthropic's cautious limited release of Mythos, citing its "strikingly capable" nature in computer security tasks. AISI's independent findings confirm Mythos's prowess, particularly in executing complex, multi-step cyberattack sequences, a significant departure from previous AI models.
While Mythos performs similarly to other leading AI models like GPT-5.4 in individual cybersecurity tasks, its true distinction lies in its ability to "chain" these tasks effectively. AISI's "The Last Ones" (TLO) test, a simulated 32-step data extraction attack on a corporate network, saw Mythos become the first model to complete the challenge from start to finish. Despite only succeeding in 3 out of 10 attempts, its average performance of 22 completed steps far surpassed competitors. AISI cautions that these simulations lack real-world active defenders and specific vulnerabilities, but the results underscore Mythos's potential to autonomously target weakly defended enterprise systems. The institute advises that as AI capabilities advance, defensive strategies must similarly leverage AI to bolster cybersecurity.
Read Also
- Morningstar's Diversified Portfolio Outperforms Traditional 60/40 Allocation
- United-American Merger: Antitrust Hurdles and Rising Airfares
- Jamie Dimon: AI Creates More Vulnerabilities for Cyberattacks
- U.S. Stocks Edge Higher as S&P 500 Nears Record Peak
- Meta and Broadcom Ink AI Chip Deal Through 2029, CEO Tan Exits Board
