Anthropic, the American artificial intelligence company behind the Claude chatbot, said it detected and disrupted efforts by foreign actors to use its technology for weapons research.

The company reported that individuals linked to state programs in restricted regions attempted to enlist its AI model in work spanning biological weapons, ballistic missiles and naval targeting. In one case described by the company, actors tied to Iran sought to use Claude in efforts to target U.S. Navy warships. In another, individuals connected to the Houthi movement attempted to use the model to help develop ballistic missiles. Anthropic also said state-linked scientists from banned regions turned to Claude for virus research.

Anthropic said it identified the activity and cut off the accounts involved. The disclosures were part of a broader accounting the company offered of how its systems have been probed by actors seeking assistance with sensitive military and biological questions.

The episode arrives as researchers warn that advanced AI models could lower the barriers to dangerous biological work. The concern is not that a chatbot invents a weapon on its own, but that it can compress specialized knowledge, guide a user through complex steps, and make expertise that was once confined to trained specialists more widely reachable. Analysts examining the field have described how the same tools that speed legitimate scientific research can also shorten the path for someone with harmful intent.

AI companies have built filters and monitoring systems intended to refuse requests tied to weapons development. Anthropic's account illustrates both the promise and the limits of that approach: the company said its safeguards flagged the activity, but the attempts themselves show that determined users continue to test those boundaries.

The disclosures have added momentum to a debate in Washington over how to regulate artificial intelligence. Lawmakers and policy researchers are weighing how much responsibility should rest with the companies that build these models, how much visibility governments should have into misuse, and what rules should govern access to the most capable systems. Safety advocates point to cases like Anthropic's as evidence that oversight is needed. Others in the debate caution against rules that could constrain a fast-moving industry central to national competitiveness.

For now, Anthropic's report offers a public window into how the company says its frontier AI models are being tested by adversaries, and into the countermeasures it says it has deployed in response. Anthropic presented the disclosures as an example of its safety systems working. Whether such systems can keep pace as the technology grows more powerful is among the questions at the center of the policy discussion.

References

1. 1.Anthropic — company disclosures on detecting and disrupting foreign actors' attempts to misuse Claude for weapons research
2. 2.News reporting — Iran-linked efforts to target U.S. Navy warships and Houthi-linked ballistic missile development attempts
3. 3.News reporting — state-linked virus research use of Claude and account terminations
4. 4.Policy analysis — Washington debate over AI regulation, corporate responsibility, and government visibility into misuse
5. 5.Research and analyst commentary — concerns that advanced AI could lower barriers to dangerous biological work