SAN FRANCISCO, September 12 — Anthropic published a report documenting five categories of misuse of its Claude AI models for military, espionage and repressive purposes, the San Francisco-based company said, drawing on an analysis of thousands of conversations flagged by its automated safety systems and human reviewers. The record shows a company founded in 2021 by former OpenAI executives, among them siblings Dario Amodei and Daniela Amodei, cataloguing how its own products were turned against civilians and governments alike.
The report was released in 2025 and arrives amid a global surge in state adoption of large language models and growing scrutiny of AI companies’ responsibility for downstream harms.
Five categories of misuse
The first category is battlefield targeting and military planning. Anthropic said it identified users in conflict zones employing Claude to generate tactical recommendations, draft military doctrine, and assist with targeting decisions. The company did not name specific armed forces, but the examples align with known efforts by militaries worldwide to integrate AI into command-and-control systems.
The second category is surveillance and espionage. According to the filing, state-linked actors used Claude to draft phishing emails, translate intercepted communications, and analyze open-source intelligence.
Such activity mirrors documented campaigns by intelligence services that have increasingly turned to commercial AI tools. Censorship and propaganda forms the third category. Anthropic found users generating disinformation narratives, automating social media posts, and drafting content for state media.
The fourth is repression of dissent, including the drafting of arrest warrants, surveillance summaries, and interrogation questions targeting activists and journalists. The fifth is cyber operations, with users seeking to write malware, identify software vulnerabilities, and plan network intrusions.
Enforcement and disclosure
Anthropic said it detected these uses through its Usage Policy enforcement and has banned accounts tied to the activity. The company has also expanded its threat intelligence team and said it shares findings with law enforcement and other AI developers. The report does not specify the number of accounts banned or the countries involved, citing security concerns.
Anthropic has not disclosed revenue figures for 2025, but its Claude models are used by millions of consumers and thousands of businesses worldwide. On the timeline, the disclosure follows earlier Anthropic reports about state-linked influence operations, including a 2024 report on networks targeting the US presidential election.
Anthropic’s findings echo concerns raised by the United Nations and human rights groups about AI’s role in enabling authoritarian practices. The company has called for stronger international norms and said it supports regulation that would require transparency from AI developers. The report underscores the dual-use dilemma facing AI companies: the same models that can boost productivity can also be repurposed for harm, and the industry is under pressure to prevent such misuse without stifling legitimate applications.
What to watch next is whether Anthropic’s call for binding transparency rules gains traction with governments now weighing AI regulation, and whether other developers follow its lead in publishing misuse findings of their own.
