Dual-use describes knowledge with both legitimate and harmful applications — the same techniques that produce a vaccine can also engineer a more dangerous pathogen. Biosafety researchers have wrestled with this for decades, but powerful AI models add a new wrinkle: they can now assist with complex scientific work well enough to sit right on that line.

In September 2026, Anthropic published a threat intelligence report detailing cases where Claude was used for biology work with weapons-relevant risk. One involved a scientist asking Claude to help draft a grant application for gain-of-function research on chikungunya virus — a request that looked like legitimate science. Anthropic acknowledged it can no longer assume its newest models fall below the threshold where they could meaningfully assist weapons development.

There's no clean solution. A blocked request might be genuine research; an approved one might not be.