19:40Machine-translated from Azerbaijani by AIRead in Azerbaijani
Iran-linked groups developed military systems with Claude
Anthropic has released its September 2026 threat report. According to the document, Iran-linked groups used the Claude artificial intelligence model to gather intelligence about U.S. warships and prepare target-selection recommendations. The Houthis also used the model to work on missile programs. The cases identified demonstrate the serious risks associated with using commercial artificial intelligence to develop weapons and support military operations.
The Iran-linked group combined transponder codes that publicly show the locations of ships and aircraft with images taken by commercial satellites. Information about U.S. Navy movements was also added. The group collected the names of U.S. military personnel from captions accompanying publicly available military photographs. It also investigated known security vulnerabilities in “Cobham Sailor VSAT” satellite communication terminals used on ships, “Cisco” equipment and “Schneider Electric EcoStruxure” systems. A group in Houthi-controlled northern Yemen used “Claude Code” to work on 3 weapons programs. One of the projects was a guided missile equipped with a smartphone-grade computer that homes in on its target during the final stage of flight. Another project involved a multistage ballistic missile capable of traveling more than 2,000 kilometers. The “R2000” missile family also included a hypersonic glide vehicle variant.
The group used the model to write guidance, navigation and control software. Open-source autopilot software was adapted for a smartphone-grade flight computer. Separate code was also developed for control and calculating the device’s position. Artificial intelligence was also involved in tuning parameters, configuring firmware and simulating flight tests on a computer. According to the report, several separate instances of the model wrote and checked code, conducted research and performed tests. In this way, they worked together like a team of programmers.
Another unit linked to Iran’s security agencies analyzed 155,216 posts over 1 year. As a result, profiles were created on 6,388 opposition figures. The collected information was used to identify and track those individuals. Another group created surveillance tools for use inside the country. One of them is a “Firefox” extension called “al-Najm al-thāqib,” which collects user IDs from major social platforms on a large scale. This extension was also used in real-world conditions. Anthropic blocked 16 Claude accounts linked to Iran’s paramilitary and domestic security agencies. However, it is not known whether the blocking covered all accounts used by those organizations. The company shared its findings with government agencies and deployed additional detection methods. The report says that groups linked to China and Russia have also used Claude artificial intelligence in military programs.
The Iran-linked group combined transponder codes that publicly show the locations of ships and aircraft with images taken by commercial satellites. Information about U.S. Navy movements was also added. The group collected the names of U.S. military personnel from captions accompanying publicly available military photographs. It also investigated known security vulnerabilities in “Cobham Sailor VSAT” satellite communication terminals used on ships, “Cisco” equipment and “Schneider Electric EcoStruxure” systems. A group in Houthi-controlled northern Yemen used “Claude Code” to work on 3 weapons programs. One of the projects was a guided missile equipped with a smartphone-grade computer that homes in on its target during the final stage of flight. Another project involved a multistage ballistic missile capable of traveling more than 2,000 kilometers. The “R2000” missile family also included a hypersonic glide vehicle variant. The group used the model to write guidance, navigation and control software. Open-source autopilot software was adapted for a smartphone-grade flight computer. Separate code was also developed for control and calculating the device’s position. Artificial intelligence was also involved in tuning parameters, configuring firmware and simulating flight tests on a computer. According to the report, several separate instances of the model wrote and checked code, conducted research and performed tests. In this way, they worked together like a team of programmers. Another unit linked to Iran’s security agencies analyzed 155,216 posts over 1 year. As a result, profiles were created on 6,388 opposition figures. The collected information was used to identify and track those individuals. Another group created surveillance tools for use inside the country. One of them is a “Firefox” extension called “al-Najm al-thāqib,” which collects user IDs from major social platforms on a large scale. This extension was also used in real-world conditions. Anthropic blocked 16 Claude accounts linked to Iran’s paramilitary and domestic security agencies. However, it is not known whether the blocking covered all accounts used by those organizations. The company shared its findings with government agencies and deployed additional detection methods. The report says that groups linked to China and Russia have also used Claude artificial intelligence in military programs.