• Anthropic documented five cases in which Claude was used for biological research that could support weapons development
  • One reseller used US infrastructure, synthetic accounts and model fallback to bypass regional and safety controls

The fact

Anthropic said it identified five cases in which Claude had been used for biological research that could support the development of biological weapons. The activity was investigated between December 2025 and August 2026 and included research involving avian influenza, orthopoxviruses, venoms, and toxins.

One of the cases involved a reseller that gave researchers in a region not served by Anthropic access to Claude through US infrastructure. According to Anthropic, the platform also relied on a zero-data-retention service, grey-market resellers and synthetic accounts. If Claude refused a request, the platform could send it to another AI model through a fallback system.

Anthropic said the reseller served more than a dozen unrelated customers and exchanged tens of thousands of messages with Claude over several days. The company later banned accounts linked to the activity and said it had strengthened its safeguards. Anthropic did not report that any of the cases resulted in a biological weapon being produced or that it had identified plans for an imminent attack.

The assessment

Anthropic's report shows that safeguards can become harder to enforce when access to an AI model passes through an intermediary. A provider may block users in certain regions or refuse a sensitive request, but a reseller can sit between the provider and the person making that request.

Routing adds another complication. In the case Anthropic described, requests rejected by Claude could be sent to another AI model. A refusal from one provider therefore did not necessarily end the user's attempt to obtain an answer. Providers also have to balance monitoring with legitimate privacy requirements. Some customers use services that retain little or no prompt data, which can reduce the information available when suspicious activity is investigated.

For BTW readers, AI safety increasingly depends on the infrastructure around the model as well as the model itself. Account identity, reseller access, API credentials, routing and logging all affect what a provider can see and stop, while those controls still have to protect legitimate customer privacy.

What to watch

Watch for changes to rules covering AI resellers, proxy services and access from unsupported regions. Tighter identity checks, API controls, model-routing policies or changes to zero-data-retention services would show how providers respond to the access gaps Anthropic identified.