Anthropic Details AI Misuse in Conventional Weapons and Intelligence Targeting

Anthropic's Threat Intelligence team reported on September 10, 2026, that they identified and disrupted operations involving malicious actors attempting to use Claude over the previous eight months. [1] This newly published report includes instances where threat actors used AI models for surveillance and the development of conventional weapons. [2] The company previously disrupted a highly sophisticated AI-orchestrated cyber espionage campaign in September 2025. [1]

Anthropic's Frontier Red Team created new evaluations to measure model capabilities in tactical intelligence targeting and conventional weapons development, such as engineering drones to strike moving targets. [2] The evaluations revealed that models can execute tasks in these domains that were historically restricted to scarce, highly-trained human experts. [2] To block such misuse, Anthropic has implemented new classifiers as on-platform safety measures. [2] Additionally, tested open-weights models from PRC developers demonstrated concerning abilities to identify adversaries and improve weapon performance, despite trailing the frontier. [2]

Anthropic’s recent disclosures underscore that AI misuse has officially expanded beyond cybersecurity and biorisk into conventional military applications and tactical intelligence. The fact that open-weights models from PRC developers also demonstrate concerning abilities in these domains indicates that the risk of weapons development and surveillance misuse is not limited to frontier platforms like Claude. Consequently, the implementation of on-platform classifiers will only address a fraction of the broader global security landscape as open models proliferate. What the sources don't address: how many distinct threat actors or nation-states were involved in the disrupted operations over the past eight months.