Palo Alto Networks CEO Nikesh Arora authorized an internal stress test using Anthropic's unreleased Claude Mythos 5, a frontier AI model designed for advanced coding and cybersecurity work. Over roughly one month, the model identified 75 vulnerabilities across more than 130 products, compared with the company's usual discovery rate of fewer than five vulnerabilities per month. Arora said the exercise exposed security gaps that would have taken five to seven years to find manually. Mythos generated working exploits in more than 70% of Palo Alto Networks' internal evaluations, drawing the attention of the US government. Around June 12, the government imposed export controls on Mythos and Anthropic's related Fable model, temporarily requiring Anthropic to disable access for all users. The controls were partially lifted around June 26 for approved entities under tighter guardrails. Anthropic had distributed Mythos through Project Glasswing, a consortium that began providing the Mythos Preview to roughly 40 or more organizations around April 7, including Palo Alto Networks and CrowdStrike. Palo Alto Networks' Unit 42 threat intelligence team used the findings to implement 26 CVE fixes, and on September 22 the company launched Unit 42 Continuous Frontier AI Defense, which combines Mythos 5 with proprietary threat intelligence for continuous exposure management.