AI Model Reportedly Blocks State-Linked Attempts to Weaponize Viruses
Anthropic disclosed that its Claude AI refused multiple requests from covert accounts seeking help to enhance viral virulence and toxin production. The company identified five particularly concerning cases between November 2025 and September 2026, with actors using anonymization and evading regional restrictions. Anthropic emphasized the difficulty of distinguishing malicious biological research from legitimate vaccine or epidemiology work.
Anthropic's internal review spanned a ten-month window, uncovering five distinct attempts to misuse its Claude model for biological weapon development. The requests focused on enhancing viral traits, specifically increasing mutation rates and virulence in chikungunya, and exploring how avian flu might adapt to mammals.
To conceal their identities, these actors employed sophisticated evasion tactics, including routing traffic through US-based third-party platforms and using gray-market resellers to bypass geo-restrictions. Despite being banned and reported to authorities, some individuals persistently attempted to regain access using zero-data-retention services.
This disclosure could heighten scrutiny on AI safety protocols, potentially prompting stricter regulatory frameworks for dual-use biological research. Legitimate scientists may face increased friction when using AI for vaccine development, while security agencies could gain new tools to track covert state-linked activities. Ultimately, the balance between fostering beneficial biotech and preventing misuse may shape the future of AI governance in high-risk fields.