Government-linked accounts have repeatedly tried to use Anthropic's Claude model for dual-use biological research that could support bioweapons development, the AI company said Thursday. The attempts, detailed in a new threat report, show state-affiliated researchers evading regional access blocks to conduct work that skirts the line between legitimate science and weapons-related applications.
Anthropic identified five cases in which users tied to government institutions bypassed its safeguards. The researchers sought to use Claude for tasks involving disease research with potential military applications - work that falls under what experts call dual-use biological research.
How the access blocks were evaded
The company did not disclose which governments were involved or name the specific institutions. In each case, the accounts used methods to hide their true locations and affiliations. Anthropic's systems later detected the activity and terminated access.
The report underscores a persistent challenge for AI companies: balancing open access to models with preventing misuse by state actors. Regional blocks and usage policies can slow bad actors, but determined users with technical resources continue to probe for gaps.
Dual-use research in the AI era
Dual-use biological research refers to scientific work that can advance medicine and public health but also be redirected toward developing biological weapons. AI models that can synthesize and explain complex scientific literature make it easier for researchers to navigate this territory - raising the stakes for companies that build those tools.
Anthropic has publicly committed to monitoring for such risks. The company's threat report marks one of the more detailed accounts an AI developer has shared about state-linked attempts to repurpose commercial models for sensitive research fields.
Why this matters for government professionals
For government security and policy teams, the report signals that AI misuse attempts are not hypothetical. State-linked researchers are actively testing the boundaries of commercial AI tools. Security protocols that rely on regional blocking alone will face persistent evasion attempts. Agencies should treat AI access controls as an ongoing operational challenge rather than a one-time compliance checkbox. The dual-use problem also sharpens the policy question around export controls and how governments classify AI model access for foreign researchers affiliated with state institutions.
Your membership also unlocks:
