September 11, 2026

Anthropic published its most detailed threat intelligence report to date, covering misuse disrupted between December 2025 and August 2026 across seven harm areas: cyber operations, surveillance, influence operations, scams, conventional weapons, biological misuse and illicit model distillation. A companion post attributes distillation campaigns to China-based labs including Moonshot AI, DeepSeek and Alibaba — Moonshot covertly routed roughly 300,000 Kimi user requests to Claude over ten days — with state-linked campaigns named from Russia, China and Iran. The report lands two days after the US NSA, CISA and FBI advisory on industrial-scale distillation.

Alongside its threat report, Anthropic's Frontier Red Team published new evaluations measuring whether AI models can perform tactical intelligence targeting, such as locating people from fragmentary information, and conventional weapons development tasks, such as engineering drones to strike moving targets. The team found models can now handle some tasks historically restricted to scarce, highly trained specialists.

The framework decouples world-state evolution from visual generation: an agent compiles natural-language rules into executable state-transition programs, a lightweight engine maintains a persistent global state (including off-screen entities), and state-augmented 3D bounding boxes are deterministically compiled into conditioning for a pretrained video renderer. On the new CombatStateBench it hits 94% count accuracy and 98% state accuracy, far above existing interactive video world models.