Anthropic sounds alarm on AI models being misused to develop biological weapons

Anthropic released its latest threat intelligence report, which addresses several real-world case studies in which scientific researchers used the company’s artificial intelligence models to develop biological weapons.The 154-page report, made public on Thursday, is considered the AI industry’s first known admission that the development of biological weapons overlaps with AI misuse.“To date, such evidence of real-world potential misuse from AI models comes from academic research, government reports, and the work of international organizations and journalists,” Anthropic said in the report. “But AI companies—whose models might be directly involved in these activities—have so far been absent from the public conversation.” “To our knowledge, no private company, AI or otherwise, has yet shared evidence of the potential misuse of their platforms for biological weapons development publicly,” it added.One of the listed case studies involved gain-of-function research on the mosquito-borne chikungunya virus. In this instance, Claude was requested to assist in writing a grant application for scientific funding related to research focused on the “virus’ transmissibility and immune evasion properties.”Anthropic’s biological safety classifier blocked the human actor’s request and launched a threat investigation into the matter following its initial review. Further investigation was warranted, the company said, because civilian researchers intended to conduct the chikungunya virus research at a military research institute in an unknown country.The identities of the scientists involved in this research or in any other biological misuse cases were not provided by Anthropic.“We are withholding the names of research institutions, the countries wherein the activity took place, and the specific biological agents or research techniques involved,” the frontier AI lab said. “The individuals implicated in these case studies are working scientists. We do not assert that they intended harm, and identifying them or their labs could expose them to harm.”Anthropic took steps to sever Claude’s involvement in said research following the threat investigation.“At the conclusion of our investigation in May 2026, we banned all associated accounts, worked with partners to take down the relay networks that evaded regional blocks, and shared our findings with affected AI labs and government authorities,” the report states.In a similar case study, Claude was used to help research avian influenza, or bird flu, and the pathogen’s “enhanced pandemic potential.”Based outside the United States, the researcher asked Claude to analyze data and parse through scientific literature, as well as assist in writing the research findings. While the research plan was in its early stages, it was serious enough to catch Anthropic’s attention this year.The compromised models in this case were Claude Sonnet 4 and Haiku 4.5, which fall below faster models like Claude Opus 5. Because the lower-class models are not accustomed to “perform expert-level biology research tasks,” Anthropic reasoned their output was limited and less beneficial for the researcher’s purposes.From this case, Anthropic concluded that the “existing safeguards on our frontier models are robust enough to force researchers to use weaker and less safeguarded models.”The Claude Mythos and Fable models are far more powerful and have stronger safeguards in place, so they’re less likely to be misused. There was only one case in which these models were misused, according to the report.Since Mythos is not available to the general public, it’s unlikely that some bad actor will get their hands on the model. Fable could potentially be misused in cyber operations, for example, because it is publicly available, but it has more robust classifiers developed by Anthropic to minimize this risk.Despite the risks that AI-assisted biological research poses, Anthropic recognizes that some scientists may want to do some good in the world by trying to find a cure for a disease. The company describes this tension between beneficial and harmful as dual use, referencing the distinction between civilian purposes and military or criminal applications.Regardless of which side the biological research falls on, the intent is not always stated. Thus, it’s difficult for AI companies to distinguish between the two.Anthropic, one of several top AI firms, released three different threat intelligence reports last year. None of them addressed biological misuse in the AI ecosystem. This appears to be a relatively new trend as AI models become more advanced.The latest report touches on other areas of concern, such as conventional weapons development and AI-augmented cyber operations.In the past year, foreign threat actors used Claude to develop software for conventional weapons like missiles or bombs. Anthropic moved to disrupt these illicit activities by banning accounts violating its usage policy and terms of service, among other actions. The company included six of these cases in its published findings: three in China, two in Russia, and one in Yemen.In one of the six cases, Russian freelancers used Claude Code to build the software system for an autonomous military drone swarm that would target enemy forces during the Russia-Ukraine war.Some Russia-based actors also employed AI to drive their cyber operations aimed at targeting members of the Ukrainian government, military, and diplomatic staff, as well as military drone technology manufacturers.Anthropic noted its concern that AI-augmented cyber operations, which are becoming increasingly autonomous, are closing the gap between state-sponsored and criminal cyber actors in terms of the sophisticated cyber arsenal available to both.The report comes two days after Jacob Coxon, an AI researcher at Anthropic, resigned because he viewed the developing technology as too dangerous. In a viral statement, Coxon openly admitted that AI developers “earnestly believe that [AI] could kill us all by the end of the decade.”Other Anthropic researchers agreed with him.EL SAYED CITES ANTHROPIC RESEARCHERS’ EXISTENTIAL WARNING IN CALL FOR AI PAUSECoxon’s departure caught the attention of Sen. Bernie Sanders (I-VT), who recently announced plans to introduce a bill in coordination with Rep. Greg Casar (D-TX) to ban artificial superintelligence and temporarily pause advanced AI development.The Ban Artificial Superintelligence Act has not been formally introduced in Congress yet. Sanders plans to hold a private briefing on Sept. 16, giving senators the opportunity to discuss AI with top researchers in the field.