Detecting and countering misuse of AI: September 2026
Detecting and countering misuse of AI: September 2026
Anthropic’s September investigation describes selected misuse of its models, including surveillance operations it attributes to government-linked users. In China, a user associated with municipal cyber police ran Claude Code with browser and database tools to gather material on a dissident and send daily reports to supervisors. Other Chinese users asked Claude for political targeting guidance. This is an operational agent workflow in the provider’s account, not just a synthetic test.
Anthropic also describes Iranian users building surveillance software, including a browser collector shipped into a case-management system. In Mali, a likely intelligence-service contractor used Claude to engineer a platform for joining telecom records and generating phone-number dossiers. About 25 million SIMs were in the reported data scope; the dossier workflow’s approval step was removed at the operator’s request. The deployed platform used other models locally, so banning the contractor’s Claude account stopped further assistance but did not take down the system. Existing government data access and human-set targeting goals supplied the consequential permissions.
These are Anthropic’s selected account and code observations, not independently audited operator or telecom logs. The SIM count does not establish how many people were dossiered, and no matched human analyst study shows whether AI increased wrongful targeting or useful leads. The Chinese police attribution has uncertainty; none of the cases establishes an AI that chose its own coercive goals. The next check is authenticated tool-call and approval records alongside comparable human work and actual subject outcomes.