Anthropic Blocks Claude Abuse for Weapon Engineering

Anthropic describes several instances where its model Claude has supported military work: drone swarms, anti-torpedo systems, missile software, and naval targeting collection. The company claims to have discovered these activities during internal investigations, closed accounts, and strengthened its safeguards. The examples involve actors in Russia, China, Yemen, and those linked to Iran.

Closed Accounts, Sharing with Authorities, and Failed Tests in Yemen

Anthropic states that it banned an account linked to a threatening actor associated with Iran, provided information to government authorities, and developed detections to prevent future abuses. In northern Yemen, the company reports that a guided missile test conducted by actors using Claude appears to have failed, with them returning to the tool a few hours later to understand the cause. According to the company, these sessions focused on guidance, navigation, and control software, with Claude also being used to execute simulations and diagnose failures.

Drone Swarms: A Russian Group Sought Lethal Autonomy

A small group of Russian actors, linked to a regional university and described as a freelance operation, used Claude to build the central software system for a drone swarm. The components included shared memory, fault-tolerant coordination, attack and observation behaviors, and return-to-base protocols, as well as terminal guidance. Anthropic claims that the software could select targets and order attacks without human intervention and that it was tested on real hardware. The company adds that the platform was designed for autonomous lethal engagement, with an embedded model capable of selecting targets and triggering detonations without human input.

Imagery from the Ukrainian Front and Nine Accounts Under Investigation

The same Russian groups are presented as specialized suppliers. They designed a tool based on training from combat images in Ukraine to differentiate between allied and enemy forces, particularly using a fixed geographic coordinate located in the Donetsk region as a strike site for demonstration. Anthropic reports having identified and examined nine accounts associated with these activities, noting that commercial VPNs were used to circumvent geographic restrictions. The accounts are said to have been created between late 2025 and early 2026, with their operations starting in May.

China: Anti-Torpedoes, Briefing on the US Navy, and Directed Energy Weapons

Anthropic identified an actor based in China who mobilized Claude to develop a proposal for the Chinese navy around an anti-torpedo weapon system. The tool was used to draft technical specifications in Chinese and produce a proposal of over 200 pages accompanied by an executive summary. According to the company, the actor compared their design to American programs using public information and refined their texts by asking Claude to act as a hostile critic. A briefing in Chinese on US Navy systems was also generated from open sources. The account was banned after being discovered during internal investigations. Anthropic also mentions another user in China relying on Claude to gather information on directed energy weapons.

Naval Targeting Linked to Iran Based on Open Sources

An actor linked to Iran used Claude to compile and analyze public information to issue targeting recommendations against US forces. The gathered corpus included a register of US personnel derived from public military photo captions, identifiers of accessible ship and aircraft transponders, scripts for querying commercial satellite images, and an inventory of sites exposing US naval movements.

Broader Context: Scope of Abuses and AI Governance

The reported cases fall within a broader scope of abuses that the company also ties to cybersecurity and research on biological weapons. Anthropic presents Claude as a cutting-edge American model that adversaries have attempted to exploit. The company explains that it conducted internal investigations to identify these activities, closed accounts, and strengthened its security measures. The publication comes amid a context where the Trump administration and Silicon Valley are advancing AI development amid safety concerns. Just days earlier, researcher Jacob Coxon had left Anthropic, denouncing the race to develop models and warning on X that some individuals building AI believe it could kill everyone by the end of the decade.