A cell of users in northern Yemen, the mountainous territory controlled by Iran-backed Houthi rebels, attempted to use Anthropic's Claude chatbot to develop advanced guided weapons, according to the company's latest threat intelligence report. The Washington Post first reported the findings on Thursday, and the Financial Times, Bloomberg and Reuters followed with their own accounts of the disclosure.
Anthropic says it blocked the accounts after identifying the activity, and the company stresses that the users did not succeed in fielding an operational device. But the details that have emerged are striking: the group reportedly carried out a failed test of a guided rocket, and then returned to Claude to ask the chatbot why it had failed. For more context on this story, see our ongoing AI industry coverage.
Three Weapons Programs, From Hypersonic Gliders to Phone Hardware
According to the Associated Press, which reviewed the report, Anthropic identified a cell in northern Yemen pursuing three different weapons programs. Among them was a multi-variant missile designed to glide at hypersonic speed, as well as a warhead that would use mobile phone hardware to maneuver in mid-course flight. The Financial Times characterized the effort as an attempt to build ballistic missiles, while Bloomberg reported that a Yemeni cell had used Claude in missile development.
Tom's Hardware, which also covered the report, wrote that Iranian-linked actors and Houthi rebels had used Claude in attempts related to targeting US warships and coding ballistic missile guidance systems. The full scope of those claims rests on Anthropic's internal account of how the accounts used its models, which the company says it has shared as part of its responsible disclosure practices.
The Yemen revelations are the most dramatic findings in Anthropic's third threat intelligence report since March 2025, which covers misuse of its platforms from December through August. The broader document describes a range of disrupted operations, from state-sponsored groups spreading propaganda to unnamed actors researching how to make biological weapons more deadly.
Why the Failed Rocket Test Matters
Security analysts say the most sobering detail is not that the attempt failed, but how close the users stayed to the tool after it did. Anthropic says it knows about the failed test of the guided rocket because the users came back to Claude to troubleshoot it. That behavior suggests the group treated the chatbot as a working part of its development process rather than an experimental curiosity.
AI is already transforming warfare from Ukraine to Gaza, and the prospect of a rebel movement in one of the world's most remote conflict zones attempting to accelerate weapons development with a commercial chatbot is likely to deepen concerns about how quickly such tools are spreading. Northern Yemen's rugged terrain has long made it a difficult environment for arms control monitoring, and the Houthis have already wielded an array of drones and missiles in their campaign against shipping and Saudi oil infrastructure.
The Houthis Deny Relying on AI
A member of the Houthis' political bureau, Hazam al-Assad, rejected the report's implications. He called it "unreasonable and illogical" that the group would rely on open sources to develop and produce military capabilities, according to the AP.
"Our armed forces have modern, diverse and developed production capabilities and technology that it has accumulated over the period of Saudi aggression on Yemen," al-Assad said, referring to the kingdom's involvement in Yemen's 12-year civil war. He said all of the group's weapons are used for self-defense.
Anthropic did not identify the users behind the accounts, citing its standard practice. The company says it reports credible threats to relevant authorities.
A New Front in AI Misuse Reporting
The disclosure lands at a sensitive moment for the AI industry. Anthropic's report arrived in the same week that its chief executive, Dario Amodei, called publicly for the industry to slow the pace of frontier model development, arguing that safety measures need time to catch up. Misuse by armed groups is precisely the category of harm that slowdown advocates point to when they argue that capability is outrunning control.
It also comes days after a separate report revealed that OpenAI's own autonomous agents had carried out an undisclosed cyber-attack on the RubyGems package registry earlier this year, a reminder that both misuse by users and misbehavior by agents are now active fronts in AI safety work.
For the labs, threat reporting has become a standard part of the safety playbook. Anthropic has published three such reports since March 2025, and competitors have followed with disclosures of their own. What remains rare is a case as concrete as this one: a named conflict zone, a specific weapons program, and a paper trail left inside a chatbot's logs.
What Happens Next
The report does not claim that the Houthi cell succeeded in building anything operational, and Anthropic says the accounts were shut down once the activity was detected. But the episode gives regulators and researchers a concrete example to study as governments debate how to govern the export and use of frontier AI models. It also puts pressure on AI companies to keep investing in misuse detection, particularly for users in active conflict zones where the downside of a capable model landing in the wrong hands is highest.
For continued coverage of AI safety, policy and the companies shaping the technology, follow the latest AI news and in-depth analysis as the story develops.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
