AI bioweapon risk: How unchecked models are amplifying bio‑security threats

AI bioweapon risk and the surge of dual‑use research

The latest disclosures from Anthropic reveal that its language model Claude was probed five times for assistance in virus manipulation, underscoring an emerging AI bioweapon risk that regulators struggle to contain.

In two incidents, scientists sought guidance on gain‑of‑function experiments aimed at enhancing viral transmissibility. Anthropic’s internal threat assessment, released Thursday, says the attempts were traced to actors with state‑level resources who masked their true intent.

“We view these cases not as isolated glitches but as evidence of organized dual‑use programs,” the report reads.

Researchers at leading institutions—from Fordham to Oxford—have warned that generative models can draft viral capsids, predict mutation pathways, and even fabricate nucleic‑acid sequences that slip past standard screening software. Such capabilities lower the entry barrier for poorly funded groups.

A recent survey by the Institute for Security and Technology found ↓ 70% of security experts believe AI will materially raise bioweapon development risk within ↓ 3 years.

Legislators are now debating a “kill‑switch” provision that would let federal agencies shut down models deemed existential threats. The Department of Commerce’s Center for AI Standards and Innovation (CAISI) is poised to draft mandatory safeguards, a move that could replace the current voluntary consultation model.

“The public deserves assurance that the most powerful models are screened for national‑security implications,” said Tom Inglesby of Johns Hopkins Center for Health Security in an interview with Reuters.

While some advocate for purpose‑built systems like AlphaFold that lack autonomous planning, others warn that any increase in AI autonomy amplifies uncertainty. Without enforceable standards, the line between beneficial science and weaponization may blur irreversibly.