SAN FRANCISCO -- A commercial intelligence company linked to Israel used Anthropic’s Claude artificial-intelligence models to build a system capable of identifying and politically profiling people in Iran and the Persian Gulf, according to a new report by Anthropic that reveals how AI can be used to turn ordinary social-media activity into intelligence.
The operation, associated with Israeli-Singaporean intelligence vendor S2T Unlocking Cyberspace, was designed to process large numbers of social-media posts and determine who the users were, where they were located, what demographic group they belonged to and whether they appeared to support or oppose their governments.
Anthropic said it detected the activity and subsequently banned the account involved.
The significance of the operation lies in what the system was designed to do. Instead of an intelligence analyst manually examining individual
social-media accounts, Claude was instructed to process batches of posts and turn them into structured intelligence.
The system categorized people into groups including urban residents, clerics, military personnel, young people, members of the diaspora and rural populations. It then assigned political leanings and confidence levels to its assessments.
In practical terms, such a system could allow an operator to move from a mass of apparently unrelated online conversations to a database of politically categorized individuals and communities.
Anthropic also identified more than 255 fabricated social-media accounts in a separate workstream. The company said this suggested that the operator was building a reservoir of fake identities that could potentially be used later.
The operation also instructed Claude to produce material in Persian, Arabic, English and German for personas designed to appear to belong to both pro-government and anti-government Iranian communities.
That distinction is important. The Anthropic report does not establish that these fake accounts were deployed in an actual campaign inside Iran, nor does it establish that Mossad directly ordered the operation.
Anthropic said it detected the activity while it was still at the pilot stage and found no evidence that the later stages of the surveillance chain were used against real targets before the account was banned.
The findings nevertheless add a significant technological dimension to Israel’s broader intelligence campaign against Iran. Anthropic said the activity independently corroborated a previous investigation by the journalism network Forbidden Stories into S2T and its surveillance capabilities.
The disclosure also comes amid renewed scrutiny of a separate Israeli effort to encourage regime change in Tehran.
Israeli broadcaster Channel 12 reported in August that Mossad chief Roman Gofman removed two senior officials associated with a plan to replace Iran’s government. One had headed the agency’s Intelligence Directorate and the other its Iran division.
Israeli and international reports said the plan involved encouraging unrest and opposition activity inside Iran.
The reported shake-up followed the failure of expectations that military pressure on Iran would rapidly produce an internal uprising. A subsequent analysis by Le Monde described the Israeli-U.S. effort to topple Iran’s government as having failed, citing the absence of expected defections and Kurdish intervention and Iran’s ability to withstand the assault.
Taken together, the developments show two different components of the campaign against Iran.
The first is conventional military pressure intended to weaken the Iranian state from outside. The second is an information and intelligence effort aimed at understanding, influencing or potentially manipulating political behavior inside the country.
AI changes the scale of that second effort. A system that can rapidly classify thousands of social-media users, identify political tendencies, map demographic groups and generate convincing material for artificial personas can perform in minutes or hours work that would previously have required teams of analysts.
That does not mean the technology can manufacture a popular uprising by itself. It does, however, provide intelligence operators with a far more powerful tool for identifying potential audiences, monitoring political sentiment and testing narratives.