Live
Anthropic says frontier AI can geolocate users and write drone guidance software
Anthropic says its strongest models performed as well as or better than an elite human proxy at locating photos. The models also wrote working control software for simulated drone tasks. The company says the results are limited by synthetic data, proxy comparisons, and weapons tests run only in simulations.
What happened
In evaluations dated September 10, Anthropic tested account linking, photo and text geolocation, simulated drone guidance, payload delivery, and navigation through GPS interference. Across 6,000 photos, Mythos Preview had a median error of 37 kilometers. The elite GeoGuessr proxy had a median error of 151 kilometers. Anthropic says Opus 5 and both Mythos models completed every simulated weapons task it set. No model solved the hardest GPS-spoofing case.
Why it matters now
These results suggest AI could reduce the amount of expert labor needed for surveillance and weapons software. Anthropic tested the weapons tasks in physics simulations, not on hardware. It used synthetic or historical data for the geolocation tests. These were company-run capability tests, not measurements of real-world effects.
Updates
What changed
Anthropic says frontier AI can geolocate users and write drone guidance software
Anthropic says its strongest models located outdoor photos with median errors of 37 to 47 kilometers. The models also wrote working flight software for simulated drone tasks. The company says the tests used human proxy comparisons, synthetic or historical data, and simulations rather than hardware.
Verification
Primary evidence before publication.
Social chatter can identify a lead. It does not authorize a HotTea live story.