Anthropic said the actors repeatedly attempted to circumvent its safeguardsAnthropic has identified a Yemen-based weapons engineering cell that used its Claude artificial intelligence model to develop guidance software for missiles and a guided rocket. In a report published Thursday, the AI company said it had disrupted six operations in China, Russia and Yemen in which actors used Claude for conventional weapons development, intelligence gathering or procurement. The Yemen case involved a group that was working on multiple missile programmes, including a multi-stage ballistic missile with a stated range of more than 2,000 km and a missile system known as the “R2000” set, which included a hypersonic glide vehicle variant.Anthropic said the actors used Claude Code “in place of human software engineers” to develop guidance, navigation and control software for flying vehicles. While the company found no evidence that the group successfully fielded an operational weapon, it said the actors did conduct a test launch of a guided rocket, which appeared to fail.How Claude was used to build missile guidance softwareAccording to Anthropic, the Yemen-based cell used Claude to integrate an open-source autopilot with a phone-class flight computer and develop the software needed to control and stabilise a flying vehicle. The work included writing control and position-estimation software, tuning control settings, running firmware builds and conducting flight simulations.The group also appeared to use multiple Claude instances simultaneously, assigning different tasks to each. One instance was used for coding, another for research and a third for reviewing code produced by the first.Anthropic said the actors repeatedly attempted to circumvent its safeguards. They concealed the intended purpose of their work, split activities across multiple sessions and used other methods to prevent individual interactions from revealing the wider weapons programme.The company said its safeguards blocked many of the requests, “but not all of them”. After the guided rocket test failed, the actors returned to Claude within hours to investigate what had gone wrong.Anthropic said it banned accounts linked to the operation and shared information with public- and private-sector partners. It also found that the actors had already developed an offline simulation toolkit that did not depend on Claude or engineering software such as MATLAB. AI tools expand access to advanced weapons workThe Yemen case was one of four conventional weapons operations described by Anthropic in its report. The other cases involved software for a drone swarm, a system designed to intercept torpedoes, and targeting software for electronic warfare and air-defence suppression.The company said the actors generally used Claude to build and refine software for weapons hardware and firmware they already had expertise in and access to. The findings suggest AI was not necessarily replacing the underlying technical infrastructure, but could reduce the amount of human engineering work required to develop it.“Sophisticated attacks no longer require sophisticated attackers,” Anthropic said in the report. “The cybersecurity skills of AI models means that AI has collased the labor and tooling gap that used to separate well-resourced, state-sponsored operations from individual operators.”The report also highlights a broader problem facing AI companies: the same capabilities that can accelerate legitimate scientific and engineering work can potentially be exploited for harmful purposes. Anthropic flags risks beyond missilesAnthropic said it had identified multiple other cases involving potentially dangerous biological research, including work related to bird flu, chikungunya, orthopoxviruses and venom toxins.Over a 30-day period, the company identified about 35 “distinct research efforts” involving potentially concerning activity. It cautioned, however, that it could not establish whether those involved intended harm or were pursuing legitimate scientific research.One case involved assistance with a grant application for gain-of-function research into the transmissibility and immune evasion of chikungunya. Another involved research into how bird flu viruses adapt to mammals and cause severe illness.Anthropic said biological misuse remained one of the “most serious risks” posed by AI models. It has also introduced new classifiers designed to detect and block activity linked to high-yield explosives and weapons development.The company said it would continue incorporating findings from its investigations into its safeguards, while sharing relevant threat intelligence with governments and other organisations. The cases underline a growing challenge for AI developers: as models become capable of handling increasingly complex technical tasks, preventing their misuse becomes more difficult even when users are not directly asking them to build a weapon.