Anthropic’s Frontier Red Team built new evaluations for AI in tactical intelligence targeting and conventional weapons development. Models now match scarce human experts on some military and intelligence tasks. Open-weights models from PRC developers lagged but still showed concerning targeting and weapons capabilities. Evaluations covered identity correlation, classification, and geolocation from photos and text. Frontier models outperformed humans in geolocating outdoor photos. Models geolocated 8% of users within 1 km using text alone. Speed tests showed Mythos Preview analyzed a 37,000-word sample in 11 minutes versus 2.5 hours for a human. Results suggest AI could reduce the labor cost of intelligence targeting. The report highlights risks of AI misuse in surveillance and weapons development. Anthropic implemented new classifiers to block such misuse on its platform.
What it unlocks
AI models can now perform tactical intelligence targeting and conventional weapons development tasks previously limited to highly-trained human experts.
- 200 identity correlation tasks
- 6,000 geolocated photos
- 1,697 users in GeoText evaluation
- 37,000 words median sample size
Sources