Anthropic Report: Iran Used US AI Models to Track US Warships
A report by Anthropic reveals that Iran utilized US-developed AI models like Claude to track US Navy ships and seek military vulnerabilities. The 154-page document outlines severe national security threats, including bioweapon research and espionage.
Iran attempted to use artificial intelligence models developed in the USA to track US Navy ships in the Middle East, according to a report by the company Anthropic, which warns of emerging threats to national security. The report's key findings were covered by The Wall Street Journal.
According to the document, an Iranian-linked group collected data from automatic identification and tracking systems for ships and aircraft, military photographs, and commercial satellite imagery to track US warships and search for vulnerabilities in their communications. Anthropic stated that it successfully thwarted the Iranian plot, which utilized the advanced Claude model.
National Security Risks and Biological Research
The American newspaper noted that this finding was one of several threats detailed in the 154-page report. The document serves as a playbook detailing how US adversaries are utilizing artificial intelligence in novel ways that pose severe threats to national security: developing new missiles, conducting aggressive disinformation campaigns, spying on dissidents, and even exploring the potential development of biological weapons.
In recent days, US President Donald Trump was asked if he fears that artificial intelligence could lead to the extinction of humanity. "No, I have no such fear," he replied, explaining that his primary concern is the possibility that the United States might not win the global race for artificial intelligence supremacy.
Simultaneously, Anthropic published data detailing real-world misuse: scientists used the Claude model for research aimed at assisting in the development of biological weapons, and entities linked to hostile nations such as Iran, Yemen, and Russia utilized it for weapons development, espionage, propaganda, and surveillance. Anthropic reported detecting five such instances of biological research in recent months, marking the first time a private AI company has published findings regarding the potential use of its platform for biological weapons development. However, the company refrained from labeling these instances as direct terrorist plots, and the names of the scientists, activists, institutions, and countries involved were not disclosed.
Evasion Tactics and Internal Warnings
According to the report, users in some cases bypassed geographical blocks and concealed their true objectives to circumvent Claude's safety mechanisms. Although tests from 2025 showed that Claude models were incapable of providing significant assistance in dangerous biological research, more advanced and potentially riskier models have since been released. Anthropic announced that it has added additional layers of defense and blocked the identified accounts.
The report was published just two days after Jacob Cookson, a 27-year-old researcher who worked at Anthropic, announced his resignation and publicly criticized the firm.
"None of the companies are operating responsibly," he wrote. "They are rushing toward a superintelligent AI system capable of self-improvement, gambling with all of our lives."
Cookson clarified that there is currently no immediate existential threat from existing models, but warned that the pace of advancement could rapidly change that. Owen Habing, who leads alignment research at Anthropic, publicly agreed and estimated that there is a greater than ten percent chance that artificial intelligence could cause human extinction within the coming decade.
