Claude Misused to Target U.S. Warships and Develop Hypersonic Weapons

Anthropic’s threat report published in September 2026 revealed misuse of Claude by Iran-linked actors. It states that U.S.-made generative AI was used to support targeting of U.S. naval forces, weapons development, domestic surveillance, and propaganda activities. Anton Shilov of Tom’s Hardware summarized the case as follows.

Iran-linked actors use Anthropic’s Claude to target U.S. naval forces, develop weapons, and spy on citizens of Iran.

The report showed that the combination of publicly available information, commercial technology, and generative AI has reached a level capable of military diversion. It is drawing attention as a case involving both defense and AI safety measures.

Tactics for Targeting the U.S.

Navy and Sources Collected

Iran-linked actors used Claude for reconnaissance of U.S. naval forces in the Middle East region and for creating targeting recommendations. They cross-referenced publicly available transponder identifiers for ships and aircraft with commercial satellite imagery, and layered U.S. Navy movement information on top. They also extracted the names of U.S. military personnel from captions of publicly released military photos.

Investigation of vulnerabilities in shipboard communications equipment was also confirmed. The targets included Cobham Sailor VSAT terminals, Cisco communications equipment, and Schneider Electric EcoStruxure systems. They examined known flaws to explore possibilities for intrusion into or disruption of communications infrastructure.

Anthropic suspended the accounts of the users involved. It introduced additional detection mechanisms and shared the findings with government authorities. Because aggregation of publicly available information itself resembles legitimate research in appearance, designing detection criteria will be a challenge.

Three Weapons Programs by a Houthi Cell

Use of Claude Code by a cell in the Houthi-controlled area of northern Yemen was also reported. The Houthis are said to receive support from Iran, raising concern about sharing of research and development results. The cell pursued three weapons programs in parallel.

The first was a guided rocket that uses a phone-grade flight computer to assist terminal guidance. The second was a multi-stage ballistic missile with a range exceeding 2,000 km. The third was the R2000 missile series, including a hypersonic glide vehicle derivative. All were programs aimed at improving long-range strike capability and guidance accuracy.

If the developers cooperate with Iran, which has an industrial base, the shift from prototype to mass production could accelerate. Sharing of design information would affect the regional military balance. The report showed the reality that even non-state actors can approach the development of advanced guided weapons.

Guidance and Control Development Support via

Claude Code

The cell made extensive use of Claude for program development for guidance, navigation, and control. It advanced integration of an open-source autopilot system with a phone-grade flight computer. It iterated on creating control laws and position estimation code, tuning variables, and building firmware. It also used generative AI to conduct flight simulations.

Notably, it used multiple Claude instances as a substitute for a group of program engineers. It divided code creation and code review, research, and simulation among them. It was an operation that compensated for personnel constraints and maintained development speed.

This method simultaneously achieves externalization of expertise and parallelization of work. It can accumulate records of trial and error and reflect them in the next design change. The fact that even a small, elite organization can sustain guidance and control development raised the threat assessment.

Diversion to Domestic Surveillance and

Propaganda and Detection Measures

The report also touched on diversions other than weapons development. It included use for surveillance activities against citizens inside Iran. It was also said to have been used to support propaganda activities.

Surveillance and propaganda share a technical foundation with targeting and weapons development. Text generation, summarization, translation, and organization of subjects can be applied directly. Military operations and information control are streamlined on the same foundation.

Development of testing infrastructure is a prerequisite for safety measures. The concept of testing infrastructure discussed in Inspired by GNOME OS Test Center and Apple TestFlight also applies to the verification process for misuse detection. A mechanism to continuously measure the effectiveness of detection systems is essential.

Current Status and Challenges of Safety

Measures and Information Sharing

Anthropic responded with suspensions, stronger detection, and sharing with government authorities. Excluding individual actors alone cannot prevent recurrence. The combination of publicly broadcast identification signals, commercial imagery, and open-source programs can be reproduced by anyone.

Satellite image interpretation depends on the quality of the display environment. The interest in reproducing creator intent shown in LG OLED evo, Creator Original Picture Mode with Prime Video is not unrelated to refining image interpretation. There is a cycle in which improvements in civilian technology increase the accuracy of military diversion.

Going forward, measures on the communications equipment side will also be important. Fixes are required for known flaws in equipment from Cobham, Cisco, and Schneider Electric. Risks must be reduced through both management of publicly available information in fleet operations and equipment updates.

Editorial Opinion

In the short term, major foundation model providers are likely to strengthen misuse detection and usage screening. Inspections and updates for known vulnerabilities in maritime communications equipment may advance. Frameworks for information sharing with government agencies are also likely to be developed at an accelerated pace.

In the long term, the lowering of barriers to weapons development through the combination of civilian AI and publicly available information is likely to change the premises of national security. Management of commercial satellite imagery, publicly broadcast identification signals, and open-source autopilot systems will likely become issues. Balancing development support functions with safety measures is likely to become the central design challenge.

The remaining question is how far support for aggregating publicly available information should be regarded as legitimate use. Questions also remain about effectiveness if detection evasion advances. How to set boundaries that curb misuse while preserving convenience is being tested.

References

• “ Iran and Houthi rebels used Anthropic’s Claude AI to target US warships and build hypersonic missiles — Houthi rebels also used the bot to code ballistic missile guidance systems “, by Anton Shilov — Tom’s Hardware, 2026-09-12T15:03:10.000Z (ARR)
• Source URL: https://www.tomshardware.com/tech-industry/artificial-intelligence/iran-and-houthi-rebels-used-anthropics-claude-ai-to-target-us-warships-and-build-hypersonic-missiles-houthi-rebels-also-used-the-bot-to-code-ballistic-missile-guidance-systems