Anthropic stated in a threat intelligence report that an entity linked to Iran utilized Claude to gather open-source information, attempting to track and suggest targets for US military vessels in the Middle East. The company interrupted this activity and banned related accounts after discovering it, while also notifying government agencies.

The report covers abuse cases identified and cut off from December 2025 to August 2026, totaling approximately 154 pages. With Claude's assistance, the entity built an automated Python pipeline that compiled a "target manual" using publicly available ship and aircraft transponder codes, personnel lists from military photo captions, commercial satellite image query scripts, and public websites exposing naval movements for identifying and tracking vessel positions. The same account also allowed the model to organize research on known vulnerabilities in onboard systems, covering maritime VSAT terminals, Cisco communication devices, and industrial control products. Anthropic did not specify the exact date of discovery and did not confirm whether these outputs were involved in actual military operations.

The report also listed other related abuses in the Middle East: a unit in northern Yemen assessed by the company as highly likely related to the Houthis developed advanced weapon guidance software using Claude, including content related to multi-stage ballistic missiles; Iranian state-aligned accounts were also used for propaganda, domestic surveillance, and creating open-source profiles of Israeli government and non-government personnel, as well as Jewish diaspora organizations. The company stated that it has banned multiple related accounts, including those associated with the Iranian Ministry of Culture and Islamic Guidance, the Khorasan Razavi Islamic Propaganda Office, and the Bina Cultural Observation Station.

Anthropic characterized these cases as unconventional abuse rather than routine use and stated that it has implemented protections and shared intelligence with public and private sectors. The report also covered attempts related to biological weapons research, weapon specifications prepared by Chinese defense manufacturers for the PLA Navy, and cyber activities related to Russia, forming a list of how adversaries can repurpose commercial models into reconnaissance, propaganda, and weapon assistance tools.

In market mechanisms, this is a security disclosure rather than a secondary market transaction. The seller is publicly accessible cutting-edge model capabilities; the buyer is state-aligned accounts automating open-source intelligence. The beneficiaries are adversaries who can use the model as an intelligence assistant until their accounts are banned; the pressured party is the US Navy's public digital footprint in the Middle East and model companies that must sever the connection between military abuse and civilian access. Funds and computing power continue to flow through commercial APIs, while governance costs are recouped by account bans, reporting, and submissions to the government.

The disclosure comes against the backdrop of attacks targeting US naval vessels in the Gulf region. The company clarified that what was cut off were accounts and pipelines, not that it claimed to have prevented a specific attack.

Source: Public Information