Scientific American: cybersecurity experts divided on Mythos severity as Bank of England intensifies AI risk testing

Scientific American published an expert analysis on 17 April assessing the cybersecurity implications of Anthropic’s Claude Mythos Preview. The 245-page system card released alongside the model describes it as operating at the level of a senior software engineer in finding and exploiting vulnerabilities, scoring 31 percentage points above the previous flagship model on advanced mathematical reasoning. The Bank of England confirmed AI risk testing had intensified following the disclosure, and German banks consulted authorities and cyber experts. The cybersecurity community is divided: former UK National Cyber Security Centre CEO Ciaran Martin said it was unlikely to prove the end of the world, while Georgia Tech’s Peter Swire noted that every CISO should take Mythos seriously even if expected harm is lower than worst-case scenarios. The UK’s AI Safety Institute tested the model and acknowledged its capabilities while noting that test conditions lacked real-world software defences. Project Glasswing involves more than 40 organisations with $100 million in usage credits and $4 million in open-source security funding.

In this domain: AI in police investigations: Longeye and Axon tools tested as courts wrestle with disclosure rules · Deepfake detection becomes an operational requirement for law enforcement as provenance methods prove insufficient · ISACA: Agentic AI adoption is outpacing security governance as enterprises deploy autonomous systems at scale