Anthropic Threat Report Exposes Attempts To Build Biological Weapons With Claude
Anthropic's detailed report outlines five specific case studies occurring between late 2025 and mid-2026 where researchers attempted to leverage Claude’s advanced capabilities for dual-use biological work. In one incident from May of this year, Anthropic's automated biological safety classifiers flagged a request asking the model to draft a grant application for gain-of-function research involving the chikungunya virus.
The proposed project aimed to genetically alter the mosquito-borne virus to increase its transmissibility and immune evasion, and was made even more concerning due to the user’s affiliation with a military research institute. When Claude initially refused the query, the user attempted to circumvent those blocks and safeguards by using VPNs to query less restricted models.
Other cases involved researchers attempting to study mammalian adaptation in highly pathogenic avian influenza (yes, the last thing we need is another COVID pandemic), as well as experiments focused on orthopoxviruses and the automated design of toxic venom peptides.
Anthropic emphasized that identifying intent in biological research is inherently complex, as legitimate efforts to create vaccines or treatments often rely on the exact same scientific methodologies needed to engineer harmful pathogens. Jacob Klein, Anthropic's head of threat intelligence, highlighted that actors are rarely asking models to explicitly build a bioweapon. Instead, the threat emerges from incremental, highly specialized assistance that lowers technical barriers for dual-use applications.
The report also detailed misuse across other severe domains, including state-sponsored malware generation, mass surveillance tools, and software development for conventional weapons like drones and missiles.
In response, Anthropic deployed strict automated classifiers, banned offending accounts, blocked regional proxy networks, and intentionally restricted suspicious accounts to older, less capable model tiers like Claude Sonnet 4 and Haiku 4.5. The company noted that its latest frontier models, such as Claude Fable 5, feature significantly tightened safeguards that reject dual-use biological queries entirely.
Anthropic chose not to publicly name the individual scientists or academic institutions involved, citing concerns that public exposure could jeopardize legitimate ongoing research efforts. However, the company confirmed that all findings have been shared with government authorities, peer AI safety organizations, and industry partners to help establish stronger collective defenses against any potential AI-assisted bioweapon development.
