Pentagon tests OpenAI and Google models to replace Claude in classified systems — Anthropic's safety-first posture may disadvantage it for military deployment
Reporting this week confirms the Pentagon is actively testing OpenAI and Google frontier models as potential replacements for Claude in classified systems. Anthropic's Project Glasswing covers defensive cybersecurity — Claude Mythos reportedly found 23,019 vulnerabilities under the program — but the safety-first posture is being read as a constraint for military-application use cases.
The procurement read is the substantive piece. Anthropic positioned Claude as the most-aligned frontier model and built its enterprise tier around that positioning; the Pentagon evaluation reads that positioning as "refuses too often" for use cases that don't have civilian analogs. OpenAI has historically been more permissive on military and dual-use applications; Google has the Gemini Robotics partnerships with Boston Dynamics and the foundation-model integration with DOD-adjacent supplier base.
The Project Glasswing carve-out gives Anthropic a defensible federal foothold even if it loses the broader classified-system tier. 23,019 vulnerabilities found in Claude Mythos defensive deployment is a substantive proof point for the AISI-aligned cybersecurity-first segment. But the binary choice the Pentagon is reportedly framing — Claude OR OpenAI/Google for classified — would be a structural revenue ceiling Anthropic has not yet had to navigate. With the Trump EO formalizing voluntary frontier access as the gateway to federal procurement, the strategic question gets sharper this quarter.
BuildFastWithAI — AI News Today June 8 2026: Pentagon AI testing → · White House — Promoting Advanced Artificial Intelligence Innovation and Security EO →