Het verhaal
On 10 Sep 2026 Anthropic’s Frontier Red Team published Measuring AI capabilities in intelligence targeting and conventional weapons — capability evals beyond cyber/bio. On tactical targeting: Mythos Preview tops account-linkage/classification on synthetic social corpora and photo geolocation (median error 37 km across 6,000 YFCC images, ~24% within 1 km), beating Champion-tier GeoGuessr human baselines; text-to-home geolocation with sandboxed search clusters Opus/Mythos around ~20 km median. On weapons: models iterate GNC software for simulated quadcopters — terminal guidance to a moving vehicle, payload drops, and GPS-denied/spoofed navigation. Opus 5 leads (e.g. 80% strike on parked high-contrast targets; hardest camouflage/evasion settings largely unsolved). PRC open-weights (Kimi K3) sit between Sonnet and Mythos on several tasks. Anthropic ties this to Threat Intelligence misuse in surveillance and drone work and says Safeguards shipped new weapons-development classifiers. Simulation-only caveats apply.