Anthropic Report Highlights Potential Misuse of AI in Biological Research

<p>Researchers outside the United States have utilized advanced AI models to plan experiments involving highly pathogenic bird flu, chikungunya, and other biological agents, according to a threat report from Anthropic published on September 10, 2026. The report indicates that these researchers employed AI to analyze viruses’ transmissibility and immune evasion, investigate orthopoxvirus immune-response genes, and design novel toxins, while attempting to bypass safeguards that restrict access to dangerous biological capabilities.</p><p>Anthropic detailed five specific cases in the report but did not disclose the countries, institutions, or specific biological agents involved. The company emphasized that it does not assert that the researchers intended to develop biological weapons.</p><p>The findings provide a tangible insight into a threat that has previously been discussed in hypothetical terms. Scientists with existing biological expertise and laboratory access are leveraging powerful AI systems to accelerate sensitive dual-use research.</p><p>These developments occur amid ongoing discussions regarding catastrophic risks associated with AI. Jacob Coxon, a former researcher at Anthropic and OpenAI, recently resigned, claiming that these companies are hastening towards self-improving superintelligence, which he believes poses significant risks. His statement garnered over 100 million views on X.</p><p>Evan Hubinger, Anthropic’s alignment science lead, expressed a personal belief that there is more than a 10% chance AI could lead to human extinction within the next decade.</p><p>Anthropic’s report highlights a current threat: researchers using advanced AI for potentially hazardous biological work, circumventing geographic restrictions, and adapting when safeguards are encountered. In one instance, Anthropic blocked a grant application for chikungunya gain-of-function research at a military research institute, which aimed to identify mutations that could enhance the virus’s transmissibility and immune evasion.</p><p>The researchers intended to engineer these mutations into infectious clones and pass the virus through live animals to select for variants that cause severe disease. While this research could aid in developing better vaccines and treatments, it also risks creating a more dangerous pathogen that could be indistinguishable from a natural outbreak, according to Anthropic.</p><p>The researchers accessed Anthropic's AI model, Claude, through a life sciences platform that circumvented U.S. infrastructure restrictions. When Claude declined to respond to sensitive prompts, the platform redirected those requests to another model with less stringent safeguards.</p><p>Despite Anthropic banning accounts linked to this activity, the platform operator quickly regained access using new identities, while Claude continued to assist in preparing research materials.</p><p>In another case, a researcher utilized Claude to plan experiments involving mammalian adaptation and airborne transmission of highly pathogenic avian influenza. The researcher studied viral traits that could enable bird flu to spread more effectively among mammals and cause severe disease outside the respiratory system. A version capable of efficient human-to-human transmission would pose a significant pandemic threat, as noted in the report.</p><p>Anthropic indicated that this researcher appeared to have access to virus isolates and animal-model facilities, confining the work to weaker models and limiting assistance primarily to study design, data analysis, and editorial support.</p><p>The report also described a separate user connected to a state-associated infectious disease laboratory who used Claude to draft a grant application for orthopoxvirus research, including a proposal to study viral genes that disable the human immune response.</p><p>Anthropic stated that the same knowledge could potentially help in preserving or enhancing the virus's immune evasion capabilities. The final two cases involved venom peptides and toxins, where one researcher built an atlas of toxin peptides and another computationally redesigned several toxins, including those associated with priority disease threats identified by the World Health Organization.</p><p>The report illustrates the challenges of preventing biological misuse through automated filters alone. While Anthropic's safeguards blocked clear attempts involving enhanced pandemic pathogens, they were less effective when users framed their work as medical or therapeutic research.</p><p>Anthropic noted that biological capabilities are dual-use, meaning the same information can be used to develop both biological weapons and beneficial medical treatments. The company recently reviewed 30 days of activity linked to adversarial state institutions and identified approximately 35 distinct biological research efforts, with some showing notable dual-use potential.</p><p>Anthropic clarified that the findings do not indicate that Claude has facilitated an imminent biological attack but do suggest that researchers associated with state programs are attempting to bypass access controls to utilize advanced AI models in biological research with potential weapons applications. The company stated that future safeguards may need to include identity verification, institutional vetting, and closer monitoring of users seeking access to advanced biological capabilities.</p>