What's this about?
Anthropic tested account linking, photo/text geolocation and simulated drone guidance. Strong models compressed work that once required scarce analysts or control engineers, including 11 minutes for a median 37,000-word account-correlation sample.
What exactly is a targeting evaluation?
It measures steps used to find and locate a person or object from scattered clues. Here the tests used licensed images, anonymized posts and synthetic social-media worlds rather than real operational targeting.
Why is it worth attention?
The photo-location result is striking and open-weight models also showed useful capability. Human comparison is only a GeoGuessr proxy, synthetic text is imperfect and drone tests cannot establish battlefield reliability.