Measuring tactical intelligence targeting and conventional weapons capabilities of AI models

Anthropic's Frontier Red Team published new evaluations measuring AI capability in tactical intelligence targeting (e.g., locating people from fragmentary information) and conventional weapons development (e.g., engineering drones to strike moving targets). The headline finding: 'For some tasks in military and intelligence domains, models could do things that, historically, only a set of scarce, highly-trained human experts could do.' Using a pipeline of 200 simulated tasks across social-media identity-correlation scenarios (WhatsApp, Telegram, Instagram, Facebook), the team found frontier models making consistent progress on 'find, fix, track, target' kill-chain tasks; open-weight models from PRC developers tested behind the frontier but still showed 'concerning ability to identify and target adversaries, and improve weapon performance.' Anthropic states this motivated new on-platform classifiers to block such misuse.

Defense and intelligence-adjacent enterprises and policymakers need to recalibrate assumptions about which military/intelligence tasks require scarce human expertise versus what current frontier and even second-tier open-weight models can now approximate.

Brief national-security and defense-sector clients on the shifting capability floor for AI-enabled targeting and weapons-development tasks when assessing platform safeguards and export-control posture.