UNIDIR warns AI and biology convergence demands layered defences against bioweapons

Anthropic's report puts misuse in focus

UNIDIR's commentary starts from Anthropic's threat report of 10 September 2026, which provided five case studies on biological misuse. According to UNIDIR, these were the first publicly documented examples of attempted biological misuse of AI models covered in detail, including the safeguards in place to disrupt such efforts.

UNIDIR is careful about what the findings show. They provide evidence that actors have already attempted to misuse AI for sensitive and potentially weaponized biological research. They do not, however, demonstrate that AI models enable the development of biological weapons, nor that such efforts would have meaningfully advanced the prospect of success. The institute reads them as a sign that actors see strategic value in using frontier AI models for dual-use biological activities that may be inaccessible through traditional pathways.

The report also describes cases in which users took evasive measures to hide their tracks and circumvent existing safety mechanisms. UNIDIR concludes that implementation strategies must be developed with varied obfuscation methods in mind.

AI is spreading through the sciences

The commentary cites bibliometric studies showing how widely AI has diffused across scientific fields. AI-related natural science publications rose to nearly 80,000 in 2025, up 26% from 2024. They accounted for roughly 6-9% of scientific output across major scientific domains in 2025, compared with less than 1% in 2010. Growth in AI-biotechnology research, including gene editing, drug discovery and bioinformatics, has been particularly rapid.

UNIDIR says models once confined to niche uses are turning into general-purpose capabilities available to specialists and non-specialists alike. Biological threats have traditionally been tied to access to specialized material, facilities and expertise. Frontier AI models, the institute says, increasingly let actors acquire and apply biological knowledge more efficiently than a few years ago, potentially lowering some knowledge and expertise barriers around sensitive dual-use research.

In its assessment, AI raises both the floor and the ceiling of what could be achieved: novices can reach previously inaccessible scientific knowledge, while experts can overcome domain-specific hurdles. A wide range of actors could also use AI to facilitate mis- and disinformation related to biological weapons. UNIDIR warns of threats beyond traditional conceptions of bioweapons, such as localized attacks, terrorism and economic sabotage through disruption of agriculture or livestock supplies.

A web of prevention and the role of industry

UNIDIR stresses that no single safeguard would be sufficient and that a layered approach is needed. It also argues that policy should address the broader convergence of emerging science and technology rather than focus on any single capability.

The central international instrument is the 1972 Biological Weapons Convention (BWC), which has 189 States Parties. It prohibits, among other things, the development, production and stockpiling of biological weapons, defined broadly enough to include any weapon developed using AI. Many States have written the prohibition into national law and added other regulatory and normative measures, helping build what UNIDIR calls a "web of prevention".

The institute welcomes frontier AI companies joining multilateral discussions on threat models, biological risk assessment and evaluations of their tools, noting that the private sector built the overwhelming majority of frontier models. But it says the most capable models continue to show the least transparency. The international community cannot rely wholly on private firms to catch every case of AI exploitation, UNIDIR adds, and the industry cannot stay detached from multilateral efforts to strengthen biological weapons controls.

A decisive moment for the BWC

UNIDIR points to clear operational limits in the Convention. It has no verification mechanism, even though States tried in the 1990s to negotiate a protocol with one. It is supported by a four-person Implementation Support Unit hosted within the United Nations Office for Disarmament Affairs, whereas treaties on chemical and nuclear weapons are backed by organizations with hundreds or even thousands of staff.

The Working Group on strengthening the BWC will hold its tenth and final session in December 2026. A draft report contains a package of measures:

• a plan for further work on verification through a dedicated open-ended working group;
• a mechanism for reviewing science and technology;
• a second mechanism focused on international cooperation and assistance.

Success would pave the way for the BWC Review Conference in 2027. Failure to reach consensus in December, UNIDIR warns, could set biological disarmament diplomacy back 25 years, to 2001, when negotiations on a protocol collapsed. Such a collapse may not irreparably damage the BWC or the current prohibitions, but it would severely curtail momentum toward strengthening the treaty.

For AI developers and governments alike, the message is that biosecurity governance of AI will be tied to how the BWC process fares, and that coordination between industry and multilateral frameworks is expected to matter.