Palo Alto Networks tests Anthropic’s Claude Mythos to enhance cybersecurity

Palo Alto Networks CEO Nikesh Arora decided to stress-test his own company’s defenses using one of the most powerful AI models ever built. The results were both impressive and a little terrifying. In April, Arora’s team turned Anthropic’s unreleased Claude Mythos 5 model loose on Palo Alto Networks’ internal infrastructure with a single directive: find vulnerabilities. Over the course of roughly one month, the model uncovered 75 vulnerabilities across more than 130 products. For context, the company’s typical discovery rate was fewer than five per month. Arora said the testing surfaced security gaps that would have taken five to seven years to find manually. A model so effective the government stepped in Claude Mythos 5 is Anthropic’s frontier AI model purpose-built for advanced coding and cybersecurity work. During Palo Alto Networks’ internal evaluations, the model successfully generated working exploits more than 70% of the time. That capability caught Washington’s attention. Around June 12, the US government imposed export controls on both Mythos and a related model called Fable, temporarily forcing Anthropic to disable access for all users. The controls were partially lifted a...