Anthropic Reports Attempts to Misuse AI for Biological Research
Anthropic says it has disrupted multiple attempts to use its Claude artificial intelligence models for advanced biological research that could have both beneficial and harmful applications. In a 154-page report titled “Detecting and Countering Misuse of AI,” the company described five cases involving work on viruses, toxins and other biological agents.
The company said it banned associated accounts, helped dismantle relay networks used to evade regional restrictions, informed government authorities and other AI laboratories, and added new safeguards. Anthropic did not accuse the researchers of intending harm. It noted that biological knowledge is often dual-use because information supporting vaccines and treatments may also enable dangerous modifications.
One case involved a request for help drafting a grant proposal concerning gain-of-function research on chikungunya, a mosquito-borne virus. Anthropic said the proposed work included changes that could increase transmissibility and severity. Claude blocked the request, but the company later found that researchers had used a third-party service to route rejected prompts to another model. Other cases concerned bird flu, orthopoxviruses, venoms and toxins.
The report also described alleged misuse linked to Iran. Anthropic said some actors used Claude for influence operations, surveillance and analysis of publicly available information about U.S. naval forces. The company said it closed the accounts, improved detection systems and shared intelligence with authorities.
Anthropic said sophisticated users may exploit the ambiguity of legitimate scientific research to avoid detection. The disclosures come amid broader debate over whether AI developers are moving quickly enough to manage emerging risks.
Separately, a senior Anthropic safety researcher recently estimated that advanced AI carries a greater than 10% chance of causing human extinction within a decade. Former employee Jacob Coxon also resigned, accusing Anthropic and OpenAI of pursuing increasingly capable systems irresponsibly. Anthropic did not immediately respond to the report’s publisher for comment on Coxon’s allegations.