Top News

Hyderabad: An Iran-linked account used Anthropic’s Claude artificial intelligence (AI) models to gather publicly available information on US naval forces and turn it into targeting recommendations, the US company said in a report published on Thursday, September 10.

The operators built what Anthropic called “targeting handbooks,” using a Python tool they wrote with the model’s help to identify and track the positions of American ships from open sources.

The material included the names of US personnel lifted from captions on public military photographs, transponder identifiers for ships and aircraft, scripts for querying commercial satellite imagery and a list of public websites that reveal US naval movements, according to the report.

One tap — then press “Add” on Google

The same account used Claude to research weaknesses in shipboard systems. It catalogued known security flaws in maritime satellite terminals, Cisco communications equipment and industrial control products, with specific vulnerability codes listed in the report.

Anthropic said it banned the account and shared what it found with government authorities. It did not say whether anything produced with Claude was used in a real Iranian military operation.

The disclosure follows a series of Iranian strikes on US naval targets in the Gulf during this year’s war. Iran’s Revolutionary Guard has said it hit the headquarters of the US Fifth Fleet at Naval Support Activity Bahrain, an attack Washington says damaged the facility without causing American deaths.

Spy tools built with AI

The report says a second set of Iranian cases involved surveillance of Iranians themselves.

Anthropic said it banned 16 accounts run by two units it assesses with high confidence were tied to Iranian paramilitary domestic security agencies. It found evidence the units were working on instructions from senior Iranian officials.

One unit, based in Qom, used the model as its engineering department. Its main product was a malicious Firefox add-on shipped to real users, disguised as a prayer-times utility, that harvested identities in bulk from social networks. The unit also built a tool that strips anonymity from messaging accounts, one that turns a phone number into a person’s identity and a fake login page for Iran’s national ID system.

Both units fed their collection into a shared case file system called “Arman.” A person’s file held their national ID number, their beliefs, their criminal record, their social accounts and a tab labelled “action.”

The second unit described itself as a seven-department body with offices across Iran’s provinces. It claimed to have surveilled and profiled 6,388 Iranians in a single year and used Claude to run analysis across 155,216 social media posts, naming 39 opposition and diaspora accounts to monitor.

Anthropic said Claude refused when the operators asked it to profile people directly but conceded that its safeguards “did not refuse many of the surveillance software tooling requests.”

Funeral and succession plans

The report also describes three accounts traced to named Iranian state propaganda bodies.

An account linked to the Islamic Culture and Communications Organization, which sits under Iran’s Ministry of Culture and Islamic Guidance, used Claude to produce full organisational plans for the funeral of Iran’s Supreme Leader and for the succession.

Ayatollah Ali Khamenei was killed in a US-Israeli strike on February 28 and buried at the Imam Reza shrine in Mashhad on July 9, after a six-day state funeral. He was succeeded by his son, Mojtaba Khamenei.

Separately, an operator working alongside one of the surveillance units used a Claude feature for saved instructions to clone the voices of three Iranian writers and prepare succession narratives in advance, the company said.

Anthropic does not offer its services in Iran. Operators reached the models using VPNs, foreign phone numbers and servers abroad.