Anthropic: Cell in northern Yemen used Claude Code to program missile guidance
In a report published Tuesday, October 13, 2026, Anthropic said an engineering cell in northern Yemen used Claude Code to develop guidance and navigation software for three missile programs before the accounts linked to the activity were suspended.
U.S. company Anthropic released a report, published Tuesday, October 13, 2026, saying it had detected cases of misuse of artificial intelligence tools, including an engineering cell in northern Yemen that used the Claude Code tool to develop guidance, navigation and control software for three weapons programs.
The report identified three programs targeted for software development: a guided missile that relies on a smartphone‑level flight computer with terminal guidance in the final stage; a multi‑stage ballistic missile for which the cell sought a range exceeding 2,000 kilometers; and a family of multi‑variant missiles known as "R2000" that includes a version equipped with a hypersonic glide vehicle.
The company said the cell ran multiple concurrent instances of Claude with specialized roles, assigning one instance to write code, another to research, and a third to review code, in an attempt to mimic a small engineering team. The tool was used to integrate an open‑source autopilot and to write control software, position estimation, tuning settings, and to run flight simulations.
Anthropic said that the tool's safety mechanisms blocked many requests but did not stop all of them, and that users attempted to split the work across multiple sessions to hide the full objective of each individual request.
The report found no evidence of a ready-to-use system being published, but it noted that the cell carried out a launch test of the guided missile that appeared to fail; within hours they again used the tool to analyze the causes of the failure, and they also developed a simulation sandbox that operates without relying on Claude.
The company said it suspended the accounts linked to this activity and shared technical intelligence with partners in the public and private sectors to help mitigate the risks.
The report highlights the gap between the rapid spread of AI programming tools and the ability of monitoring systems to detect misuse when the work is broken into tasks that appear innocuous on their own.
