Anthropic says its AI models were misused in attempts linked to biological weapons research

Anthropic said it has disrupted several attempts to use its artificial intelligence models for research that could help develop biological weapons. The company said the cases were identified over the past eight months and were set out in a report published on Thursday. It said the misuse involved its Claude chatbot and other models, and that it acted cautiously where it could not determine whether the work was legitimate or malicious.

Sponsored

One case in May involved a scientist seeking help with a grant application to engineer mutations to chikungunya, a mosquito-borne virus, in a way that would make it more harmful in live animals. Anthropic said the case was concerning in part because it appeared to be linked to a military research institute. The company said it banned the accounts involved.

Jacob Klein, Anthropic's head of threat intelligence, said the firm could not know whether the research was intended to be weaponised. The report also said the company had seen attempts to use its models for missile projects and espionage, alongside what it described as state-sponsored surveillance uses. Anthropic said evaluations of its older models last year found they could not meaningfully assist dangerous biological research, but that newer systems can complete more complex scientific work.

That shift has sharpened concern among researchers and policy experts about how advanced AI tools might lower the technical barrier to harmful work. The findings come amid a wider debate over how to balance access to powerful AI systems with safeguards against misuse. Andrew Weber of the Council on Strategic Risks, who reviewed the report before publication, said access to such models should be limited to trusted researchers.

Sponsored

Anthropic said the challenge is that the same kind of inquiry that can support vaccines can also be used to help engineer dangerous pathogens, making intent difficult to judge in real time. The company's warning adds to growing scrutiny of how frontier AI models are being used beyond their intended purposes. It also follows a separate report from the firm on surveillance operations and model abuse, suggesting that misuse is not limited to one type of actor or one region.

The broader concern is that as models become more capable, they may be used to accelerate work that would previously have required specialist expertise and larger teams. What remains unclear is how far any of the flagged biological research progressed, whether any of the attempts were connected to formal state programmes, and whether other companies are seeing similar patterns. Anthropic said it had blocked the accounts involved, but it did not say whether any enforcement action followed.

The next question is whether regulators or AI firms will move toward tighter access controls for sensitive research use.

Anthropic said it has disrupted multiple state-sponsored surveillance operations that used its artificial intelligence models, with cases linked to China, Iran and west Africa. The company said the incidents were identified and blocked between January and July. It said the campaigns were aimed at dissidents, diaspora communities and minority groups.

Sponsored

According to the company, the surveillance efforts targeted pro-democracy figures in Hong Kong, Tibetan and Falun Gong communities across Asia, and Iranian minority communities and opponents of the Iranian government abroad. In one case, Iranian actors developed a method to identify people through their social media accounts. In another, a contractor working for Malian national security authorities used Claude to design the underlying software for intelligence gathering.

Anthropic said the cases show how AI can be used to support surveillance work that would previously have required a larger engineering team. The company also said it had disrupted attempts to use its models to design weapons, conduct questionable biological research and create dating scams. Separately, it accused Chinese developers of using Claude to generate responses for their own users while also distilling that output to improve their own models.

The findings add to wider concerns about how advanced AI systems can be misused by state-linked actors. Surveillance of dissidents and minority communities has long been a feature of digital repression, but the company said AI is now lowering the technical barrier for such operations. That raises questions for AI firms about how to detect abuse, limit access and respond when their tools are used in ways that may support intelligence activity.

The report also points to a broader competition around AI development and model training. Anthropic alleged that Chinese developers, including Moonshot and DeepSeek, secretly used Claude to generate answers for users, with one case involving almost 300,000 customer requests over a 10-day period. The company said those requests were routed through a network of 5,380 fraudulent accounts, most of them appearing to be located in Singapore and Japan.

It said some of the data may have contained sensitive user information. It remains unclear how many people were affected by the surveillance operations or whether any arrests or other enforcement action followed the disruptions. The company did not say whether the actors behind the campaigns were acting directly for governments or through contractors and aligned groups in every case.

What is now being watched is whether other AI firms report similar misuse patterns and whether regulators respond with tighter controls on model access and monitoring.

#Anthropic #artificialintelligence #biosecurity #chikungunya #biologicalweapons