Top science bodies now warn that today’s artificial intelligence can already help bad actors plan biological threats, even as they agree it cannot yet design a pandemic virus from scratch.
Story Snapshot
- National Academies and international reviews flag real AI-enabled bio risks, despite current limits.
- Experts say models can troubleshoot and lower barriers to harmful biology work.
- Security think tanks push “defense in depth” instead of waiting for disaster.
- Evidence still shows no end-to-end AI design of a human pandemic pathogen.
What The Top Reviews Actually Say
The National Academies’ 2025 review says to monitor “capability uplift” where artificial intelligence tools could help design infectious agents, model spread, or boost automated labs. The same report sets a clear limit. It says current tools cannot design and build a new transmissible agent with epidemic or pandemic potential. That means the risk is not zero, but the strongest claim is about assistance, not full creation. This nuance often gets lost in public debate.
The 2026 International Artificial Intelligence Safety Report goes further on how help can look today. It finds general-purpose systems can give detailed steps for biological and chemical weapon work and help users fix problems as they go. It also notes biological foundation models can generate designs for novel pathogens, citing a study in bacteria viruses. That is not a human threat by itself, but it shows fast-moving design capacity that deserves close watch.
Where The Line Is Drawn Today
Review authors and lab leaders agree on a key boundary. No public evidence shows artificial intelligence can now create a human pandemic virus end-to-end. The National Academies say no available tool can de novo design a novel virus, and that larger outbreaks remain unlikely from today’s tools alone. Analysts caution that help with instructions is not the same as clearing real wet-lab hurdles. Those hurdles include synthesis, testing, and controlled facilities.
At the same time, the absence of a turnkey pathway does not erase risk growth. The international report warns that models can help people bypass technical blocks and suggest workarounds for rules. That guidance can save time and reduce the need for expert mentors. In the wrong hands, even small gains matter. That is why experts focus on monitoring capabilities, screening requests, and tracking how cloud labs and gene synthesis firms filter orders.
Defense-in-Depth, Not One Silver Bullet
Policy groups urge layered safeguards instead of betting on a single fix. A 2026 analysis from the RAND Corporation argues that no single safeguard can stop an actor who tries to use artificial intelligence to build a biological weapon. RAND outlines nine steps that work together, from better model evaluations to stronger screening and faster response plans. The Center for a New American Security calls for clear testing of foundation models and tighter checks by cloud labs and DNA providers.
These plans share a theme that resonates across party lines. They do not wait for a crisis headline. They aim to measure what models can do, reveal where real danger grows, and close gaps before harm spreads. That approach reflects lessons from the last pandemic. It also answers public frustration with a government that often acts late, blames others, and protects insiders. Measurable standards and transparent testing can rebuild trust if leaders follow through.
Bridging Benefits and Risks Without Spin
The same tools that raise alarms also power drug discovery and faster vaccines. That is why the National Academies call for careful monitoring, not a halt to progress. The goal is to keep the upside while blocking misuse. Leaders must invest in independent red-team tests, publish safety results that inform the public, and set clear rules for when model capabilities cross danger lines. Those steps help citizens judge claims without hype or dismissal.
Same pattern as pandemic response. Warnings existed early. What was missing was a rehearsed response, practiced before the deadline to act arrived. Healthcare and finance run drills for exactly this reason. AI risk still has no equivalent tested playbook.
— Anand Sharma (@AnandNSharma) September 11, 2026
Several evidence gaps still need proof. Independent labs should test whether model outputs help in real wet-lab work, under strict biosafety controls. Agencies and firms should release sanitized records on flagged synthesis orders. Model makers should disclose results of biological safety tests. These actions would show where artificial intelligence truly lowers barriers, where it only adds convenience, and where it must be constrained. Until then, prevention remains the safest bet.
Sources:
theatlantic.com, nature.com, rand.org, frontiersin.org
