Until recently, imagining artificial intelligence participating in a military operation meant thinking about a system controlling a drone, identifying a target, or even deciding where to shoot. The case that Anthropic has just revealed is less spectacular, but perhaps more disturbing: an AI that takes thousands of scattered pieces of information and turns them into a dossier that can help a military force find and study its adversary.
In its intelligence report, the company claims to have detected and dismantled an operation in which an actor linked to Iran used Claude to collect and analyze public information with the aim of obtaining intelligence on US naval forces deployed in the region. Anthropic identifies the group as GTG-30005 and describes it as an actor Iran-nexusan expression that is not equivalent to a direct attribution to the Government of Iran.
Anthropic’s description is quite explicit. The group used Claude to “develop targeting recommendations against U.S. naval forces in the region.” To do this, he developed, with the help of the model itself, programs written in Python to automate a chain of tasks. In this case a software intended to identify and track naval positions.
The interesting thing is that the basic data did not necessarily come from secret military systems. Much of it was publicly available. The material collected included ship and aircraft transponder identifiers, military photographs, names of US personnel extracted from the captions of those images, queries for commercial satellite images, and a list of websites that published US ship movements. Claude helped gather, organize and analyze all that information.
The difference with respect to a conventional search is precisely there. An isolated piece of information can say little. A transponder identifier, a photograph taken in a port or a satellite image may seem like unrelated pieces. But a system capable of processing large amounts of information can establish connections between them and turn them into a much more complete image of the movements of a naval force.
The operation was also not limited to locating ships. According to Anthropic, The group also asked Claude to collect information on vulnerabilities of systems installed on board, including satellite communications terminals.Cisco communications equipment and products used in industrial control systems. The report makes it clear that Anthropic discovered the activity and ended up blocking the account. The company also developed new detection systems and shared intelligence information with government authorities to help disrupt the operation.
The episode is part of a broader phenomenon that worries Anthropic. The report analyzes operations detected between December 2025 and August 2026 in which Claude was used for cyber espionage, surveillance, propaganda, conventional weapons development, fraud and other malicious operations. The company itself recognizes that the cases described represent “the most notable and novel threat activity” that it has identified so far.
And there is one particularly significant detail. Anthropic claims that its models are not being used solely as assistants who answer questions. In some cases they are becoming a layer that connects different tasks: collecting information, writing programs, processing data and producing results that can then be used by external people or systems. In the Iranian case, Claude did not sink any ships or control any missiles. Made something simpler: It helped reduce the work needed to build military intelligence from information that was already available. Anthropic notes that “as models become more capable, their risks will also increase. Developers and governments will have to simultaneously improve their protection systems.”