Anthropic surveillance report: five activities explained

Anthropic published a report on September 10 alleging that actors used Claude to analyze surveillance material, build personal dossiers and develop surveillance software. Its main disclosures are listed below; all are based on the company’s investigation.

1. China-linked activity: identifying people for “control.” Operators analyzed the political sensitivity of social posts and news, queried existing surveillance databases and prepared daily briefings. The material concerned questioning and monitoring people’s movements and communications. The report also said a religious-affairs intelligence unit used an AI assistant to produce thousands of investigative reports a month.

2. Syria: profiling Uyghurs and attempting to recruit sources. Operators used group-chat material to build personal profiles and Claude to translate outreach messages. Targets included Uyghurs and associated armed formations. Anthropic assessed the operation as China-linked, but had low confidence that the operator was a contractor and did not confirm successful recruitment.

3. Iran and the Gulf: classifying social media users by political leaning. A commercial project Anthropic linked to S2T organized users by location, demographic category and political views, then produced briefings. It was still at the pilot stage when detected; subsequent operations against real targets were not confirmed.

4. Iran: compiling identity records and developing collection tools. Files contained national identification numbers, beliefs, criminal records and social accounts. Operators also used Claude to develop a malicious browser extension that collected social media identities. Anthropic said the extension had been deployed.

5. Mali: developing a surveillance platform using mobile-operator data. A consultant used Claude to help develop a system that generated personal dossiers for state intelligence authorities. Claude assisted with software design; the platform subsequently ran locally using other models. This does not mean all communications data was uploaded to Claude. Banning the developer’s account did not stop the local system.

This account draws on Anthropic’s report and International Business Times’ coverage. Producing dossiers, attempting contact and taking operational action are distinct stages; a tool’s reported capabilities should not all be treated as surveillance that has already occurred.