Nvidia, Palantir, and Booz Allen restrict Anthropic Claude models over autonomous cyber capabilities

Major defense contractors and AI firms are limiting deployments of Anthropic's Claude models following research revealing unprecedented autonomous hacking abilities, intensifying a showdown between the AI company's safety principles and government demands for unrestricted access. Nvidia, Palantir Technologies, and Booz Allen Hamilton have begun restricting or eliminating use of Anthropic's Claude models in their organizations, citing concerns about the models' ability to autonomously exploit real-world software vulnerabilities. The restrictions stem from a September 2026 Cyber Weapon Index assessment by Booz Allen that tested 18 leading AI models and found only one—Anthropic's Claude Mythos—capable of executing a complete autonomous cyber attack chain against enterprise networks.