State-linked bad actors in China and Russia are actively misusing Anthropic's Claude AI system for influence operations and other potentially harmful activities, according to a report published by Politico. The findings add a new layer of urgency to an ongoing debate about how frontier AI companies can prevent their technology from being turned against democratic institutions, even as those same companies race to stay competitive globally.
What the Report Says
Politico's reporting outlines how operatives tied to both Chinese and Russian networks have been using Claude in ways that violate Anthropic's usage policies. While the specific tactics vary, the pattern fits a broader trend of state actors probing and exploiting Western AI tools for disinformation, cyberoperations, and propaganda generation. Anthropic has not publicly disputed the core findings. The company has previously acknowledged that, like other frontier AI providers, it faces ongoing attempts by malicious actors to circumvent its safety systems.
Key Facts
- State-linked actors in China and Russia are reportedly using Claude for influence operations.
- The activity violates Anthropic's published acceptable use policies.
- Anthropic has safety and trust teams dedicated to detecting and stopping policy violations.
- The report comes as US-China AI competition intensifies across both commercial and national security dimensions.
- Other major AI providers, including OpenAI and Google, have faced similar misuse reports in recent months.
The timing is notable. Anthropic has been vocal about the geopolitical stakes of the AI race, and the company's researchers have published detailed scenario analyses examining what happens if Chinese AI development outpaces American efforts. A separate piece of analysis on how Anthropic maps the US-China AI race through 2028 underscores just how seriously the company takes the competitive and security dimensions of this moment. Being caught in the middle, with its own tools turned against Western interests, complicates that posture significantly.
The challenge is that any sufficiently capable AI system becomes a target. Restricting access too aggressively cedes the market; restricting it too loosely invites exactly this kind of misuse.AI security analyst, quoted in Politico
A Familiar Problem With New Stakes
Misuse of AI tools by state actors is not a new phenomenon, but the scale and sophistication are growing. Claude's expanding capabilities make it a more attractive target. Claude's model family now spans a range of performance tiers, with the most capable versions able to generate persuasive text, assist with code, and reason through complex tasks at a level that makes them genuinely useful for both legitimate users and bad actors alike.
The situation is further complicated by geography. Claude is officially unavailable in China, a restriction that has created market openings for domestic Chinese competitors. Coverage of the Claude blackout in China and its effect on rival models shows how access restrictions can reshape competitive dynamics without necessarily solving the underlying security problem. Bad actors rarely respect official distribution limits.
Anthropic's trust and safety teams are understood to be actively working to detect and terminate accounts engaged in policy-violating behavior. The company uses a combination of automated detection and human review to identify misuse. Still, the volume of API usage globally makes complete prevention difficult, and sophisticated state actors have resources to obscure their activities behind layers of intermediaries and third-party access points.
The broader policy question now landing in Washington is whether AI companies need tighter coordination with intelligence agencies to flag state-sponsored misuse in real time. Several lawmakers have raised this issue in recent months, and the Politico report is likely to give those discussions fresh momentum. For Anthropic, a company that has built much of its public identity around safety-focused AI development, the news presents a genuine reputational and operational challenge that technical guardrails alone may not be enough to resolve.
