TL;DR: Anthropic said users in Houthi-held northern Yemen used Claude Code to pursue three guided-weapons programs, and the company banned the accounts after an investigation covering December 2025 to August 2026.

Anthropic disclosed on Sept. 11, 2026, that it had detected and disrupted work in northern Yemen to use its Claude models for advanced missile development, according to The Associated Press. The company described the users as operating in territory controlled by Iran-backed Houthi rebels, and said the effort centered on guidance, navigation and control software for rockets and missiles.

The disclosure came in Anthropic's threat-intelligence publication titled Detecting and Countering Misuse of AI: September 2026, according to Al Jazeera. Anthropic said the report covers activity it disrupted between December 2025 and August 2026 across seven harm areas, including cyber operations, influence operations, surveillance, scams and fraud, biological misuse, conventional weapons development, and distillation. The company said Claude Haiku, Sonnet and Opus models were used in the cases examined, according to the company's report summary.

What Anthropic says happened in northern Yemen

According to The Associated Press, Anthropic said Claude users in northern Yemen tried to use the AI model to develop advanced missiles. The Associated Press identified northern Yemen as territory controlled by Iran-backed Houthi rebels.

According to The National, Anthropic investigated and disrupted threat actors which used its AI software for military purposes, including in Houthi-held territory in Yemen. The National reported that the group was working on three sophisticated missile projects simultaneously.

Bloomberg reported on Sept. 11, 2026, that Anthropic uncovered a group in northern Yemen — where Iran-backed Houthi militants operate — using its Claude AI model to support development of missile and rocket systems. Bloomberg reported that the unidentified actors were working on three weapons projects.

The Washington Post reported on Sept. 11, 2026, that militants in Yemen sought to use a coding tool to produce guided rockets and missiles, and that Anthropic was able to block some but not all the requests. The Washington Post said the threat report was released Thursday and identified militants in northern Yemen as a stronghold for Houthi rebels using its systems to run three weapons-development programs.

The three programs described in the report

According to The National, the projects included proposed missile variants such as a hypersonic glide vehicle, technology normally associated with the world's most advanced militaries.

One project involved turning commercially available technology into a guided rocket, using an off-the-shelf, phone-grade flight computer and a homing system designed to steer it toward its target during the final phase of flight, according to The National. A second project was described as a multistage ballistic missile that the group said it wanted to send across more than 2,000 kilometers, according to The National and Bloomberg. Bloomberg reported a ballistic missile with a claimed range of more than 2,000 kilometers.

The third and most ambitious effort was called the R2000 programme, a family of proposed missile variants that included a hypersonic glide vehicle, according to The National. The National described the hypersonic glide vehicle as a technology normally associated with some of the world's most advanced military powers.

Al Jazeera reported that the effort included a guided rocket and a long-range ballistic missile, and described the company's account as an intervention in northern Yemen to block an effort to deploy Claude for missile guidance software.

How the group used Claude Code instead of engineers

According to The National, rather than hire software engineers, the actors used Claude Code to write the guidance, navigation and control software that keeps a vehicle steady and on course. Al Jazeera similarly reported that the operators used Claude in place of human software engineers, assigning different instances of the model specific roles to write missile-guidance and flight-control software.

The National reported that several Claude instances ran in parallel with assigned roles: one wrote code, one did research and a third reviewed what the first produced, with the humans acting as the leads, handing out tasks. The Washington Post reported that the individuals used Claude's coding capabilities in the effort to produce location-guidance software for weapons.

According to Anthropic's report summary, the cases shared in the September 2026 report are not typical misuse, but rather examples of the most notable and novel threat activity identified to date. The company said it publishes the work because it believes it has a responsibility to disclose malicious misuse of its services, and said that as models become increasingly capable, their risks will increase unless AI developers and society's defenders act to make them safer.

Anthropic also said the threat actors covered in the report include suspected state-sponsored groups, financially motivated criminals, commercial spyware vendors, state propaganda institutions, and politically motivated individuals, according to the company's report summary. The company said the cases range from a network of fake dating apps designed to defraud users to surveillance systems built to identify and monitor dissidents.

Safeguards, refusals and evasion tactics

According to The National, Anthropic's safeguards refused many of these requests, though not all. The Washington Post reported that Anthropic noted it was able to block some but not all the requests.

According to The National, the group worked around safeguards by concealing what they were building and what the software was for, and by breaking the work into separate sessions so no single conversation exposed the whole picture. Al Jazeera reported that the operators avoided detection by obscuring their ultimate goals and breaking tasks across separate sessions, so no single prompt gave away the operation.

According to Anthropic's report summary, sophisticated and persistent threat actors continuously test safeguards and try to circumvent the technical measures used to detect and prevent misuse. The company said it will continue to evolve safeguards and coordinate with partners to improve its ability to detect, disrupt and prevent future misuse.

The company's report summary said that in each case covered between December 2025 and August 2026, it disrupted the activity, used what it learned to strengthen safeguards, and shared intelligence with authorities and industry partners where appropriate. Al Jazeera reported that the company banned the accounts involved and shared threat information with public- and private-sector partners to mitigate risks posed by the actors.

Test-fire, disruption and surviving capability

According to The National, the actors never got to a working weapon, as far as Anthropic could tell, but they did test-fire a guided rocket. The National reported that the test appeared to have failed, and that within hours the actors returned to Claude trying to diagnose what went wrong. Al Jazeera reported that the company said it has no evidence the group managed to field a working weapon, though Anthropic claimed the operators appeared to have conducted an unsuccessful test-fire.

The National summarized the timeline and response this way: the activity took place between December 2025 and August 2026, involved Claude Haiku, Sonnet and Opus models, and was disrupted by Anthropic, which banned the accounts and published the case in the report. The Washington Post similarly reported that the activity involved location-guidance software work identified in the threat report released Thursday.

According to The National, Anthropic found the activity through its own investigations into suspected weapons development, banned the accounts, and passed threat information to government and industry partners. However, The National reported that the group had already built an offline simulation toolkit that runs without Claude or commercial tools like MATLAB, so some of its capabilities survived the ban.

Anthropic's report summary said none of the misuse cases involved the use of Claude Fable or Mythos-class models, with the exception of one illicit distillation case. The summary said no malicious activity was found on Claude Fable or Mythos in the cyber-operations cases examined.

Context around Yemen and AI in warfare

The Associated Press reported that artificial intelligence is already transforming warfare from Ukraine to Gaza, and that its use on a rugged and remote battlefield is likely to increase concerns about its rapid spread.

According to The National, Iran-backed Houthis use the capital Sanaa and northern Yemen, which they control, as a launch pad for attacks against neighbors. The National reported that on Thursday, the rebels seized the strategic port city of Mokha on the western coast, further expanding their control along the Red Sea and near the Bab Al Mandeb shipping route.

The National reported that over the past decade, the Houthis have slowly transformed from a local armed faction into a wider force, threatening neighboring countries and attacking shipping lanes while consolidating control over Sanaa and north-west Yemen. The publication cited the U.S.-based Middle East Institute for an estimated troop strength of 350,000, including child soldiers, forced-conscript African refugees and foreign mercenaries.

The National also reported that the rebels rely primarily on cross-border smuggling networks and complex maritime supply routes that extend across the Arabian Sea, the Gulf of Aden and the Red Sea to secure a significant portion of their military and logistical needs, including weapons, drone components, ballistic missiles, and technology and equipment related to naval capabilities.

The company's report summary said it hopes the findings will help other developers recognize similar patterns on their own platforms, give governments and civil society a clearer view of how emerging threats take shape, and strengthen collective defenses.

Sources referenced