AI’s Next Security Problem May Be Biological

The most credible scientific bodies now warn that artificial intelligence can already lower key barriers to building biological weapons—even as they say it cannot yet design a pandemic virus end to end.

Story Snapshot

Top scientific reviews urge monitoring of artificial intelligence that can aid pathogen design.
Defense think tanks call for layered safeguards, not one-off fixes.
International analysis says artificial intelligence can provide actionable weaponization help today.
Consensus holds that full de novo pandemic-virus design is not currently possible.

What leading reviews actually say about the risk

The National Academies mapped where artificial intelligence meaningfully shifts bio-risk now and where it could soon. The review flags models that predict how infectious a pathogen might be, suggest genetic edits, and coordinate automated labs as areas to watch. It stops short of claiming today’s tools can build a pandemic virus from scratch. It urges close tracking of capability “uplift,” including design of fully replicating agents and de novo virus design as red lines to monitor.

That measured line matters. It grounds the debate in what these systems can do this year, not in science fiction. It also rejects the easy out of “nothing to see here.” When a national academy says the tools cannot yet design a novel pandemic pathogen but can already accelerate steps around design and testing, the adult response is prudent guardrails. That is the same playbook we use for aviation, nuclear plants, and water safety: layer defenses early while the risk curve is still bendable.

How artificial intelligence changes the attacker’s playbook

The 2026 International AI Safety Report states that general-purpose models can already give detailed, step-by-step information relevant to biological and chemical weapons. It adds that these systems help users troubleshoot and bypass practical hurdles, from protocols to procurement. Dedicated biological models go further. They can generate novel genome designs in constrained cases, like altered bacteriophages, which shows direction of travel even if it is not a human pathogen. That mix shifts time, talent, and cost curves.

RAND’s analysis cuts through the hype. It states no single safeguard can stop a determined actor from using artificial intelligence in a biological attack and lays out a defense-in-depth plan. The strategy stacks nine interventions across model governance, cloud labs, DNA synthesis screening, and incident response. That approach mirrors counterterrorism lessons since 2001: force an adversary to clear many hurdles, catch more attempts early, and shrink the blast radius if one slips through.

What the hard limits are today—and why they should not lull us

The same National Academies review draws a bright line: today’s tools cannot de novo design and then build a transmissible pandemic-capable agent. They also say no current artificial intelligence tool can design a completely new virus. The report estimates that, for now, the scale of harm from misuse would most likely be local, not global. Those caveats are real and should shape how we talk about risk without melodrama or fatalism.

Conservative common sense says use these limits to act, not to stall. When brakes work, you still install seat belts and airbags. The prudent path is targeted. Screen orders at DNA providers. Audit cloud labs. Test frontier models for biological knowledge and shut down dangerous outputs. Demand incident logs and safety evaluations from model vendors. If the risk is not yet global, that is an opportunity window to keep it that way through rules that protect both innovation and public safety.

Signals of urgency across institutions

National security groups have moved from “if” to “how to prevent.” The Center for a New American Security urges screening of foundation models, cloud labs, and synthesis providers. Their reports highlight risks of bioterrorism, engineered superviruses, and targeted bioweapons while calling for evaluation standards before deployment. That stance matches the medical rule to “first, do no harm” in the digital realm: measure, verify, and only then scale. Waiting for proof in blood is not a policy.

Public voices have raised the alarm as well. Bill Gates has warned that artificial intelligence could enable bioweapons and that criteria to judge these models are still missing. Elite concern does not prove capability, but it shifts accountability. When leaders with resources and reach ask for clearer rules, lawmakers should insist on enforceable benchmarks and transparent testing, rather than glossy trust-us pledges. Industry can help write the standards. It should not be allowed to write the exemptions.

What to do before the window closes

Congress and agencies should commission red-team trials that test whether models actually reduce expertise needed for design, optimization, or troubleshooting in controlled settings. Require held-out benchmark tests for predicting transmissibility and virulence. Subpoena or mandate disclosure of flagged synthesis orders and cloud-lab incidents. Make model safety evaluations public by default, with sensitive details redacted. These steps align with mainstream reviews and keep the public debate rooted in demonstrated capability, not vibes.

Here is the bottom line. Artificial intelligence is not conjuring a pandemic in one click. It is already greasing the gears around the edges that matter. The science says monitor, test, and layer defenses now. That is the grown-up center: neither panic nor denial, but early action that keeps risk local, rare, and reversible. If we do that, the only “next pandemic” story artificial intelligence will write is how we dodged one.