Anthropic reported that it stopped several attempts by scientists to use Claude for research that could support biological weapons development. The company described five cases in which users concealed their purpose or circumvented safety controls, including activity associated with regions where Anthropic does not offer its models.
One case involved a researcher who spent weeks planning experiments related to avian influenza. Anthropic said its filters limited the work to less capable models and later banned the accounts, but it could not determine whether the researchers intended harm. Biological information is often dual use, meaning the same methods can support vaccine development or dangerous pathogen research.
The report identified biological misuse as one of the most serious potential risks from frontier AI. It also documented unrelated abuse involving fraudulent dating services, surveillance of dissidents, and attempts by other laboratories to extract model capabilities. Anthropic called for cooperation among AI companies and governments as evasion methods become more sophisticated.
Read nextAnthropic study finds AI agents sway analysts’ judgment callsScientists argue physical barriers limit risks
Scientists interviewed separately by WIRED argued that access to information is not the main obstacle to producing a biological weapon. Research scientist David Bellamy said AI can help users locate and combine protocols more quickly, but bad actors would still need genetic material, specialized equipment, virology expertise, and repeated physical experiments to assemble and test a viable pathogen.
Ginkgo Bioworks chief executive Jason Kelly said current AI systems cannot simply take control of a laboratory. Even highly automated facilities depend on humans to supply materials, operate equipment, and approve sensitive steps. Immunologist Derya Unutmaz added that defensive AI systems could also help researchers develop vaccines if a new pathogen emerged.
Other experts warned against treating those barriers as permanent. Institute for Progress fellow Olivia Scharfman said fully autonomous biological laboratories do not yet exist, but an AI system or human attacker could potentially recruit people to perform physical work. RAND biosecurity researcher Steph Guerra said uncertainty supports layered safeguards rather than reliance on any single model filter.
Experts recommend screening synthetic DNA orders
Recommended protections include mandatory screening of synthetic DNA and RNA orders, stronger controls on actionable biological instructions, better outbreak surveillance, and information sharing among model developers, gene-synthesis companies, and governments. The emerging picture is more measured than an immediate extinction scenario: AI may lower some informational barriers, but practical misuse still requires human intent, laboratory access, materials, and expertise.
