Anthropic publishes hard numbers on how well its own models can guide drone strikes and geolocate people
A new Frontier Red Team report finds Opus 5 hit an 80% terminal-guidance success rate against parked vehicles in simulation, and out-geolocates champion GeoGuessr players. Anthropic published it as a safety disclosure alongside new misuse-detection classifiers.
Anthropic's Frontier Red Team published a capability-evaluation report on September 10, 2026, measuring how well current models perform at intelligence targeting and simulated weapons guidance. It tests what the models could do if misused; it does not document misuse. On photo geolocation, Anthropic's Mythos Preview model reached a 37.0 km median error, more precise than the top 0.01% of human GeoGuessr players (151 km median error). In a simulated drone-guidance environment, Opus 5 hit an 80% terminal-guidance success rate against a parked, high-visibility vehicle, 47% against a moving target, and near zero against camouflaged or evasive targets. In a payload-delivery test against a moving target in wind, Opus 5 succeeded 28% of the time.
Why Anthropic published it
The report says the evaluations show why "on-platform safety measures are necessary." Anthropic paired it with new classifiers built to detect and block weapons-development requests, and says those classifiers will be imperfect and need iteration.
How to read it
These are simulation results under controlled test conditions. They are not evidence of drone strikes or surveillance campaigns. Anthropic doesn't prohibit military use of Claude outright; it has built detection tooling. For enterprise buyers, the percentages matter less than the posture: a frontier lab is quantifying its own models' dual-use capability in public and building mitigations in response.