Anthropic Says Claude Is Becoming a Tool for War and Repression

Anthropic Says Claude Is Becoming a Tool for War and Repression
Anthropic’s Sept. 10 threat report put a blunt proposition at the centre of the AI safety debate: harmful uses are not a distant hypothetical. The company said it had disrupted accounts linked to actors in China, Iran and elsewhere, then strengthened its safeguards.

Among the most consequential cases, Anthropic said a China-based defence researcher with links to military research institutions used Claude to build a 16-module Chinese-language suite for electronic warfare and air-defence suppression. Midway through the project, the user shifted its scenario to 12 targets in Taiwan, including radar sites, air bases and command facilities. Anthropic stressed that the work was “scenario-fed, not live ISR,” and said there was no evidence the software had been deployed by the PLA.

The report also described a separate China-linked effort to develop an anti-torpedo system for a prospective PLA Navy acquisition proposal. In that case, Claude helped draft technical specifications, review designs and build elements of fire-control software; Anthropic said it could not attribute the activity to a specific person or entity before banning the account.

Elsewhere, an Iran-linked group allegedly used Claude to assemble open-source handbooks for identifying and tracking US Navy positions in the Middle East. Anthropic said it shared its findings with authorities and added detections to reduce repeat misuse. A northern Yemeni team, meanwhile, used Claude Code in work on rocket and missile guidance software, returning to the model after an apparent guided-rocket test failure to troubleshoot the problem.

China-linked operations were also accused of using the model for domestic-security-style work: identifying Uyghurs in Syria for potential recruitment and compiling dossiers on religious leaders, dissidents and diaspora groups. One operation generated 2,475 completed reports in 30 days, according to Anthropic’s account.

The company’s intervention is central to its case for safety controls. Yet one episode offers the opposing warning: when Claude refused sensitive research aimed at modifying chikungunya, the users reportedly moved to another model with weaker safeguards.

Continue reading https://foxvector.com/stories/01a09800-a687-11d6-721d-191f89a634cb