AI can synthesize viruses; what's missing is robots with dexterous hands
Monrad believes model intelligence is nearly sufficient, but turning designs into live viruses still requires precise wet lab work—which depends on robots with dexterous hands, and current robots aren't dexterous enough.
The video won't play here. Listen to the audio instead:
The argument · timestamps estimated from transcript position
The final bottleneck is robots, not artificial intelligence
Monrad divides biosecurity risk into four stages: historical biological weapons, AI enhancing human capability (which he believes is now underway), AI coercing humans (using model-based deception or coercion to make humans complete physical tasks—he views emerging social engineering waves in other domains as leading indicators), and fully autonomous end-to-end scenarios with no human involvement. On the last category, his judgment is specific: models will soon be "smart enough," but there are simply "no thumbs dexterous enough." He frames this not as an intelligence problem but a robotics problem—though he acknowledges the situation will change by the 2030s.
— Joshua MonradHundreds of thousands of people can today reconstruct transmissible influenza
Genome length is a rough proxy for difficulty: smallpox is about 186,000 base pairs, polio about 9,000, Ebola about 20,000 to 30,000. Activating assembled genomes into live viruses presents another hurdle. But following known reverse genetics protocols, Monrad estimates hundreds of thousands worldwide could today reconstruct transmissible influenza. The real question is whether AI will expand this pool by orders of magnitude by teaching more people to complete lab procedures—tasks that once required a PhD final year now might take just weeks of lab training.
— Joshua MonradModest uplift study results lag behind actual model capability
The 2023-24 uplift studies only tested multiple-choice questions, criticized for not capturing real wet lab conditions. A later study—co-sponsored by Active Site and the Frontier Model Forum, with Sentinel also investing—cost over a million dollars and had 150 relative novices complete cell culture, molecular cloning, and virus production tasks in a wet lab, half using state-of-the-art models and half using only internet research. Results showed modest improvements in some tasks but no clear effect on end-to-end processes. Monrad reads this not as reassuring: these studies are slow and expensive; there should be ten such institutions running one quarterly, not annually—otherwise results always lag model iteration speed.
— Joshua MonradTacit knowledge barriers are being actively dismantled by technology
Sonia Ben Ouagrham-Gormley's book on biological weapons barriers once corrected early AI-biosecurity fears that underestimated wet lab know-how. But Monrad says the tide has turned, making people too complacent. These barriers aren't naturally disappearing—"we're actively dismantling them": advances in multimodal systems and computer vision, plus companies deliberately collecting human lab-work videos to train robots or smart glasses that real-time correct pipette angle. This democratizes biology, but Monrad is direct: for many tacit knowledge barriers, it's a death knell.
— Joshua MonradAgents can replicate viruses but cannot design entirely new ones
Facing the thought experiment of "misaligned agents deliberately making destructive pathogens," Monrad is specific. For well-studied viruses like influenza, data and protocols are rich—they'll perform well, reaffirming his prior prediction about Active Site results: modest improvements, which he believes will be the last such result. For genuinely novel designs, they'll struggle because biological dimensionality is too high; no dataset can simulate it at once, though he admits we've been "surprised repeatedly." For end-to-end paths requiring zero human involvement, at the time of recording in October 2026, current robotics can't produce a living pathogenic virus no matter how many aligned agents help—but he's hearing more claims that robots may be more of an "LLM-solvable problem" than thought.
— Joshua MonradDefenses work by stacking weak signals, not perfect single filters
Sentinel assessed hundreds of control candidates—cell lines, animal models, pipettes, biosafety cabinets—using two criteria: signal-to-noise ratio and concentration (scattered vs. widely available or 3D-printable). Synthetic DNA scores well on both: sequence itself reveals intent, and only ~200 companies globally can synthesize gene-length double-stranded DNA. Most candidate nodes can only be controlled by "prohibition," which would harm both defenders and all life science, making it infeasible. His answer: don't expect one perfect filter. Stack weak signals—an order that diverges from known sequences and serves unclear purpose, plus a flagged model query, plus a suspicious cloud lab application. Individually, none are red flags; together, actionable. KYC evolves into KYA: Know Your Agent.
— Joshua MonradIf I had the button, I would press it
Nathan proposes provocation: frontier model companies controlling trillions of dollars in compute should face concrete biological resilience standards before training continues—a binding "if-then" commitment. Monrad directly agrees. He says he's more confident diagnosing where the world stands on the path to "biological resilience" as a terminal state than what to pause specifically. He flags a common misjudgment: seeing some biodefense investment then thinking training can proceed unchecked—those people haven't calculated what true resilience requires, a decade-plus engineering effort. Subset goals like critical infrastructure or key personnel may move faster.
— Joshua MonradDesktop DNA synthesizers remain completely unregulated
Monrad believes social incentives diffuse responsibility—for conference chitchat and podcast soundbites on every topic, few go deep on one issue. His team has seen ordering synthetic DNA improve in recent years, with platforms more likely screening and refusing suspicious orders. But this exposes the harder frontier: desktop DNA synthesizers themselves, which need hardware security, sales-end KYC, even geolocation and trade-in buyback plans. Many have mentioned it in papers or tweets; almost no one wakes up daily to solve it. It's truly an unclaimed problem—Sentinel is hiring for it. He also highlights another neglected direction: international diplomacy, hoping Geneva, Beijing, and Singapore's top talent will treat biotech governance as the highest-value work.
— Joshua MonradIn their own words · checked verbatim
It's soon probably the case that they're smart enough, but they really don't have sufficiently opposable thumbs was sort of my take.
Joshua Monrad9:09
What might have been previously feasible by, you know, someone in the last year of their PhD now could be done by someone with a few weeks of laboratory training.
Joshua Monrad25:29
that's really amazing, but there should be 10 active sites.
Joshua Monrad32:07
It's gonna be a death knell to many of these tacit knowledge barriers.
Joshua Monrad37:15
no amount of agentic intelligence would be enough to produce a live physical virus with the current level of, like, robotics.
Joshua Monrad50:12
KYC will become KYA and it's sort of know your agent.
Joshua Monrad55:46
So you're not denying access to a model. You're not denying access to a material. You're just aggregating the information.
Joshua Monrad58:48
If I were holding that button, I would absolutely press it.
Joshua Monrad1:16:07
Figures
| Active Site uplift study participants | 150 | 28:48 |
| Global companies capable of gene-length double-stranded DNA synthesis | ~200 | 58:48 |
| US critical personnel equippable with high-performance reusable respirators for $1 billion | ~30 million | 1:22:23 |
Glossary
- uplift study
- Tests whether AI enables practitioners to complete wet lab biology tasks of greater difficulty than they could without it.
- KYA (Know Your Agent)
- An extension of Know Your Customer regulatory principles to AI agents, using stacked weak signals rather than single perfect filters to detect misuse.
- tacit knowledge
- Hands-on experimental skills and intuition that are difficult to write down and can only be learned through practice.
- managed access regime
- A governance approach that sets approval thresholds for high-risk pathogen data instead of releasing it publicly or blocking it entirely.
- reverse genetics protocol
- A laboratory procedure that reconstructs transmissible live viruses by assembling a designed or known genome.
How to listen
Founders, investors, and security researchers tracking AI's risk spillover into the physical world who want concrete mechanisms, not abstract risk discourse.
The opening framing around the OpenFace Twitter debate can be skipped.