Anthropic tests AI for intelligence targeting

Anthropic’s new red-team evaluation reports that Mythos Preview placed 23.7% of 6,000 geolocation images within 1 km, while warning that frontier and open-weight models can assist intelligence targeting and conventional-weapons work. In its capability-evaluation report, Anthropic says the tests are simulated, open-weight systems trail the frontier but still show concerning capability, and the findings prompted new misuse-detection classifiers. Builders should treat model access, monitoring, and incident response as deployment controls; a gap below the frontier is not a safety boundary, as the paper’s shrinking security lead time shows.