Anthropic has published a report on its efforts to detect and prevent the misuse of its models. The report examined cyber, influence, surveillance, fraud and weapons development activities conducted between December 2025 and August 2026 by groups believed to be state-backed, criminals and propaganda organizations; “Claude Haiku,” “Sonnet” and “Opus” were used most frequently.
The company reported that Claude was used by three actors in China, two in Russia and one in Yemen in the software of conventional weapons, as well as in intelligence and procurement processes. A group in northern Yemen used Claude Code for guidance, navigation, stabilization and simulation work on guided rockets and missiles; no evidence was found that it had developed operational weapons, and its accounts were shut down. Another China-based actor was reportedly using Claude in the fire-control software and test plan for an anti-torpedo system, and its activities were assessed to be linked to a defense manufacturer.
The cyber examples included Russian espionage, China-linked network reconnaissance, financially motivated crimes, and the targeting of political organizations and the media in France. An account monitoring Iran and the Persian Gulf was also banned in June 2026.
Why it matters
The report shows that AI tools can be used not only for content generation but also in high-risk areas such as the software of military systems, intelligence, cyber operations, and fraud. This picture reveals that, alongside state-linked entities and criminal groups, activities related to defense manufacturers also fall within model providers’ oversight. Although the closure of accounts shows that Anthropic is seeking to limit misuse not only by detecting it but also by cutting off access, the report does not conclude that all of the activities examined were prevented. The key question that remains open is how much of similar use can be detected and, particularly in weapons development and intelligence processes, within what limits the use of models can be monitored.
Background
Anthropic is not a new name in the FikirPilot archive: we have published 24 news reports mentioning the name in the last 90 days; the most recent is dated 20 September 2026.