Anthropic Releases Evaluation Results on AI Models' Tactical Intelligence Targeting and Conventional Weapons Development Capabilities

Anthropic's Frontier Red Team released a report evaluating AI models' tactical intelligence targeting and conventional weapons development capabilities. This evaluation was conducted to measure the risks of AI being misused for military and intelligence operations and to establish the basis for introducing classifiers to block such misuse.

Tactical Intelligence Targeting and Geolocation Estimation

AI models demonstrated high accuracy in targeting tasks involving identifying and connecting individuals using synthetic SNS data. In particular, the Mythos Preview model recorded performance closest to the theoretical upper bound, completing tasks that take human experts 2.5 hours in an average of 11 minutes. In geolocation estimation evaluations, Mythos Preview recorded an average error of 37km in Flickr image-based tests, surpassing the top human performers (GeoGuessr Champion 151km). In text-based location estimation, some models identified locations with accuracy within 1km, confirming potential privacy violations.

Drone Guidance and Weapons Development Capabilities

AI models demonstrated the ability to write code for FPV terminal guidance and bomb drop control using only camera and IMU data, without GPS or laser guidance. The Opus 5 model showed the highest performance, recording an 80% hit rate on stationary vehicles and a 47% hit rate on moving vehicles. Additionally, it wrote navigation code utilizing IMU dead-reckoning techniques to reach destinations with a 15–20m error margin even under GPS jamming or spoofing attack scenarios. This suggests that AI can remove software bottlenecks in precision weapons development and increase accessibility.

Implications and Response Measures

The evaluation results indicated that both closed and open-weight models can assist threat actors in human identification and weapons subsystem design. Anthropic emphasized the necessity of safety measures, such as introducing classifiers to block weapons development, to prevent such misuse. Furthermore, it recommended that democratic nations update existing laws and regulations to align with the AI era and research the potential for AI utilization from a defender's perspective.