Anthropic\'s security-focused AI model Mythos Preview reportedly found security vulnerabilities in U.S. government classified systems within hours during a controlled test, drawing industry attention. A U.S. senator said the NSA director reported the model penetrated nearly all classified systems in hours, though an anonymous official said it did not take over systems but identified weaknesses. Anthropic limits access, citing its strong ability to find and exploit software flaws.

Growing concern is emerging over a so-called patch gap, as AI can find vulnerabilities and write attack code within hours, leaving corporate security teams with less time to fix flaws. A Google M-Trends 2026 report cites cases where attacks occur before patches are deployed. Tests of Anthropic’s Mythos preview also showed rapid exploit generation for Windows kernel and Firefox vulnerabilities, while experts propose new prioritisation methods for patching.