Anthropic: Claude helped route blocked grant prompts to rival

Anthropic reported that its AI model Claude was used to help reroute a grant application it had blocked, sending the refused prompts to a competitor’s model. The company described the case in a new threat intelligence report released this month.

The blocked proposal sought funding to study chikungunya, a mosquito-borne virus that causes prolonged pain and has no cure. Civilian researchers drafted the application and listed a military institute as the host. Anthropic’s safety filter refused the submission, and investigators say the same operators then had Claude assist in writing code that forwarded the blocked prompts to another model as a workaround for the refusal.

The company published a 154-page report that documents attempts to misuse Claude for cyberattacks, influence operations, surveillance, biological research and weapons development. During a 30-day sweep, investigators identified 35 research efforts tied to state-affiliated institutions and highlighted five cases involving biology. The report includes an example in which a user drafted a smallpox-family grant in about an hour using a different model, Opus 5. Anthropic says it disrupted operations it judged harmful and blocked or banned accounts involved.

The report does not assert malicious intent by the researchers. “We do not assert that they intended harm, and identifying them or their labs could expose them to harm,” the document states. The company also noted limits to automated controls, writing, “A classifier cannot simultaneously enable benefit and prevent harm.” A separate public post from the company added, “We disrupted every operation in the report,….” The report also notes cases where blocked requests moved to less-restricted models; Anthropic described one bird flu example as largely clerical. It reported other instances where clients used its tools to track dissidents and said it cut off access when necessary.

The disclosures come after U.S. lawmakers introduced the AI Kill Switch Act in July. The bill, filed by Representatives Ted Lieu and Nathaniel Moran, would require developers to keep the ability to shut down systems. Anthropic said it is withholding lab names, has banned offending accounts and continues to monitor and block misuse while calling for additional technical and policy work to address risks.