Anthropic disrupted an AI-assisted weapons development cell in northern Yemen that used Claude Code to build guidance software for three separate weapons programs. The threat actors operated between December 2025 and August 2026, dividing tasks across multiple chat sessions to bypass automated safety guardrails.
The artificial intelligence company revealed the disruption in its September threat report covering the detection and countering of AI misuse. The report outlines how threat actors in northern Yemen bypassed safeguards by concealing their ultimate objectives, splitting tasks across multiple sessions, and assigning specific engineering roles to different instances of the AI model. Anthropic says it has shut down a cell of threat actors based in northern Yemen that used Claude to build guidance software for three weapons programmes running in parallel.
Weapons Programs and AI-Assisted Engineering
The cell worked on three distinct weapons programs simultaneously. These initiatives included a guided rocket, a multi-stage ballistic missile designed with a stated range goal exceeding 2,000 kilometers, and a multi-variant missile set designated as the R2000, which incorporated a hypersonic glide vehicle. In its September threat report on detecting and countering misuse of AI, the company says the actors used Claude Code in place of human software engineers.
Yemen terrorist group used Claude instead of software engineers
Rather than seeking simple advice on isolated engineering roadblocks, the operators ran several simultaneous instances of Claude Code to mimic a small engineering team. One instance generated code, a second conducted research, and a third reviewed the technical output. The setup is the part that stands out. The cell did not simply ask an AI model for advice on a hard problem. It ran several Claude instances at the same time and assigned each one a role, the way a lead delegates work on a small engineering team. One instance wrote the code. A second handled research. A third reviewed what the first one produced. Claude was used to write the software that steers and stabilises a flying vehicle. The model was tasked with guidance, navigation, and control—the complex software responsible for stabilizing and steering a projectile.
In one specific implementation, the operators utilized Claude to integrate an open-source autopilot onto commodity hardware resembling consumer phone-class flight computers, the sort of commodity hardware found in consumer devices. Claude wrote the control and position estimation software, tuned the control settings, ran the firmware build pipeline and performed a flight simulation. The reporting also documents the use of reinforcement learning for tuning flight controls and running six degrees of freedom trajectory simulation on the ballistic missile side.
Post-Test Diagnostics and Evasion Tactics
While Anthropic’s safety filters blocked many of the incoming requests, the actors evaded detection by masking their goals and distributing tasks so that no individual session revealed the full scope of the weapons programs. Anthropic’s safeguards blocked many of the requests. Not all of them. The actors hid their goals and concealed what the software was meant for, then split tasks across many sessions so no single one revealed the full programme.
The cell proceeded to conduct a live field test involving a guided rocket. When that test-fire appeared to fail, the operators returned to Claude within hours, uploading post-test telemetry data to diagnose the malfunction. A failed rocket test sent Yemen terrorist group back to Claude within hours. Anthropic says it has no evidence the cell managed to field an operational device. It did test-fire a guided rocket, and that field test appears to have failed. The company knows this because the actors came back to Claude within hours to work out why, bringing post-test telemetry in for diagnosis.
Yemen group used Claude for missile work
Northern Yemen is controlled by the Houthi rebels, who have been attacking shipping in the Red Sea and pushing toward the coast near the Bab el-Mandeb Strait. Responding to the findings, a member of the Houthi political bureau, Hazam al-Assad, called it unreasonable and illogical to suggest the group would rely on open sources for military production, saying its forces have their own accumulated capabilities.
Platform Bans and Persistent Toolkits
Following its review, Anthropic banned every account it could link to the cell and shared threat information with public and private sector partners. Banning the Yemen cell’s Claude accounts did not take the toolkit away. The complication is what the group had already finished.
Before the platform access was revoked, the group compiled its simulation toolkit into a standalone executable file—a digital model of the weapons system that runs without Claude and without engineering environments such as MATLAB. This allowed segments of the project to proceed with no platform access at all.
The Yemen case is one of six conventional weapons operations detailed in the report, which covers activity Anthropic disrupted between December 2025 and August 2026. Historically, intelligence agencies and United Nations panels have pieced together such programmes from recovered hardware, long after the fact. In this instance, the software was visible while it was still being written.
- Diablo Animated Series Officially in Development at Netflix
- TikTok Rejects Meta’s Ads Urging Youth Protection Measures
- Anthropic CEO Calls to Slow Down AI Development as Musk and Altman Show Support (newsdirectory3.com)
- Anthropic CEO Dario Amodei Urges Slowing AI Development Amid Safety Fears (time.news)
