AI Biosecurity Fears Meet A Reality Check

No, artificial intelligence cannot yet design a virus capable of sparking the next pandemic, but the country’s top scientists say that door is starting to creak open.

Quick Take

A 2025 National Academies review found AI tools cannot currently design a new virus able to spread as an epidemic, but says the capability gap is closing and needs constant monitoring.
RAND and the Center for a New American Security both call AI-enabled bioweapons a real prevention problem, not a hypothetical one, and recommend layered safeguards now.
Bill Gates has publicly warned that oversight of these AI models is “completely missing,” adding elite-level pressure to the debate.
Leading biosecurity experts like Kevin Esvelt and Gigi Gronvall say the actual near-term risk of an AI-driven pandemic remains low.
The real fight isn’t whether AI can build a virus today. It’s whether America moves fast enough on safeguards before it can.

What The Nation’s Top Scientific Body Actually Found

The National Academies of Sciences released a major report in April 2025 studying whether artificial intelligence could help someone build a dangerous pathogen. Researchers looked at three specific dangers: designing harmful proteins and toxins, boosting the virulence of existing pathogens, and building an entirely new virus from scratch. Their conclusion was blunt. Current AI tools cannot yet do that last, most dangerous task.

That finding matters because it directly addresses the loudest fear in this debate: a rogue actor typing a prompt and getting back blueprints for the next pandemic. The report says the science underneath viral behavior is still too poorly understood, even by AI systems, to pull that off. But the same report insists this gap won’t last forever and demands active monitoring of “capability uplift,” meaning any jump in what these tools can predict or design.

Why Serious Defense Groups Aren’t Waiting Around

The RAND Corporation isn’t a group prone to hysteria, and in August 2026 it laid out nine separate interventions to guard against AI-assisted bioweapons, describing the threat in terms usually reserved for nuclear proliferation. Their message: no single fix works, so multiple layers of defense are needed now, before capability catches up to concern. That’s not fear-mongering. That’s the same “defense in depth” thinking that protects power grids and water systems.

The Center for a New American Security reached a similar verdict in 2024, warning that AI could eventually help create “unprecedented superviruses” or targeted bioweapons aimed at specific populations. Their fix involves screening the cloud labs and genetic synthesis companies that could unknowingly manufacture a dangerous design. That’s a practical, enforceable idea, not science fiction, and it deserves congressional attention rather than being shelved as alarmist.

What AI Can Already Do, According To Global Researchers

The International AI Safety Report, a major 2026 global assessment, found that general-purpose AI systems can already hand out detailed information relevant to biological weapons and help troubleshoot technical roadblocks. That’s a lower bar than “designing a pandemic virus,” but it’s not nothing. Removing friction for a bad actor, even without full capability, is exactly the kind of quiet risk that tends to get ignored until it’s too late.

The same report points to a 2026 study where a biological AI model generated a heavily modified virus entirely from scratch. The catch: that virus infects bacteria, not humans. It’s a meaningful technical leap, but it’s not proof anyone can yet engineer a human pandemic pathogen. Treating that bacteriophage result as equivalent to a human bioweapon breakthrough would be dishonest, and credit goes to the researchers who were careful not to overstate it.

The Skeptics Have a Point Worth Taking Seriously

Kevin Esvelt, a respected biosecurity researcher, told The Atlantic the odds of an AI-driven pandemic remain “quite low”. Gigi Gronvall added that AI still has no way to actually build and test new pathogens in a lab. Georgia Adamson’s analysis for a policy think tank found experts agree the bottleneck isn’t the AI’s imagination, it’s a lack of quality viral training data. These aren’t dismissals. They’re honest boundaries on what’s provable right now.

Bill Gates And The Elite Warning Problem

Bill Gates has repeatedly warned that AI oversight for bioweapons risk is dangerously thin, saying the criteria to review these models are “completely missing”. When a tech billionaire who built his fortune on software progress starts sounding the alarm, that’s worth more attention than another activist making the same claim. It doesn’t prove an attack is coming. It does prove the people closest to the technology aren’t sleeping easy.

The honest picture sits between two extremes. AI hasn’t built a pandemic virus, and every credible technical review says it can’t yet. But the same reviews, from RAND to the National Academies to international safety researchers, agree the trend line points toward more risk, not less. Waiting for definitive proof before building safeguards would be the same mistake made with past technologies where regulation arrived only after the damage was done.