The Cognitive Revolution

Joshua Monrad on the biodefense gap AI is widening

Joshua Monrad· Co-founder and Managing Director at Sentinel Bio
·~104 min·English·Turpentine Network
AI SafetyPolicyAgentsMultimodal
TL;DR

AI will make engineered pandemics possible long before the world can defend against them, so Monrad argues the window to harden biosecurity - and maybe to pause - is closing fast.

01Core Mental Model

Four Phases, Not One Threat

Monrad splits AI bio-risk into four phases - humans alone, humans using AI, AI using humans, and fully autonomous - and says each needs its own defenses and its own timeline.

I think we can try and put dates on each of these four phases.

— Joshua Monrad, The Cognitive Revolution
Key Insight
Collapsing these into one scary number hides where the leverage is: the uplift phase is here now and demands model-level and supply-chain controls, while the fully autonomous phase is a later, different problem.
02The Real Constraint

It Is Robotics, Not Intelligence

The thing stopping a model from making a virus end-to-end today is not how smart it is - it is that it has no hands in the physical lab.

it's soon probably the case that they're smart enough, but they really don't have sufficiently opposable thumbs

— Joshua Monrad, The Cognitive Revolution
Key Insight
This reframes the near-term fight: if intelligence is already enough, the binding constraints are robotics and access to physical materials - which is exactly where Monrad argues defense should concentrate.
03Intent

Why The Omnicidal Might Choose Biology

Most would-be attackers have no use for a bioweapon - it is slow, hard, and kills indiscriminately - so the real worry is the rare actor whose goal is mass death itself.

if you look at groups that are motivated by something a lot more sinister like actual omnicide or ending life

— Joshua Monrad, The Cognitive Revolution
Key Insight
Because that intent is rare, difficulty is currently doing the protective work - Monrad notes a real anti-life attack used an LLM but drove a car bomb instead of pursuing biology, probably because biology was still too hard.
04Measurement

Uplift Is A Lagging Indicator

The studies that measure how much AI helps a novice in the wet lab are slow and expensive, so they report yesterday's models - and Monrad expects the next round to finally show real danger.

there's not a lot of uplift here and I also think this will be the last one

— Joshua Monrad, The Cognitive Revolution
Key Insight
The tacit-knowledge barrier that reassured people is being actively dismantled - multimodal models, lab-trained smart glasses, and companies collecting fine-grained lab video are all lowering the bar that the studies were measuring.
05The Dual-Use Trap

The Defender's Prompt Is The Attacker's

The same question that lets a defender find and patch a weakness is what an attacker asks to exploit it, which is why you cannot simply switch dangerous biology off.

a lot of questions and a lot of ways of working with biology are just very intensely dual use.

— Joshua Monrad, The Cognitive Revolution
Key Insight
Dual use is why refusal alone fails: filtering pathogen data from an open-weights model helps, but if that data is cheap to obtain, a small fine-tune puts the capability back - so the fight moves to materials and services, not just the model.
06Defense in Depth

You Cannot Win At The AI Layer

Since you cannot fully control the model, Monrad's team hardens the physical world instead, and found that DNA synthesis is the rare choke point worth governing.

it's relatively high signal to noise in the sense that the sequence tells you what it is.

— Joshua Monrad, The Cognitive Revolution
Key Insight
After reviewing hundreds of candidate choke points, synthetic DNA stood out on two axes - the sequence itself reveals intent, and only about 200 companies make gene-length DNA - so a few suppliers can be governed without banning the tools science needs.
07Theory of Victory

The 80 Percent Solution

You do not need every supplier to cooperate - Monrad argues that getting most DNA makers and cloud labs to screen orders already buys a major improvement in safety.

north of 80% of providers of certain materials and services such as synthetic DNA or cloud labs are both doing the kind of KYC that would allow them to refuse an illegitimate uh query

— Joshua Monrad, The Cognitive Revolution
Key Insight
This comes from working backward from a defined end state rather than chasing perfection - and it reframes screening plus cross-service signal aggregation as a tractable, fundable goal instead of an all-or-nothing mandate.
08The Hard Call

He Would Press The Pause Button

A self-described techno-optimist, Monrad still says that if he controlled the switch he would pause frontier AI for biology until defenses are built, because danger arrives long before protection does.

if I were uh holding that button, I would absolutely press it.

— Joshua Monrad, The Cognitive Revolution
Key Insight
The force of this is that it comes from someone whose life's work is the upside of biology - his point is not that the risk is certain, but that the timelines do not line up, and pretending a few grants make the world resilient is the actual fantasy.