Why DNA Screening Is Failing Generative Biology

Chris J. Murphy

Most biosecurity experts think that DNA synthesis screening will catch dangerous biological designs before they ever reach a wet lab. They are completely wrong.

Lachlan Reed

Wait, CJ, come on. Every major DNA provider cross checks every order against a black list of known pathogens, right? Like, if someone tries to print smallpox, the alarm sounds.

Chris J. Murphy

That was true when people were copying and pasting existing viruses from nature, Lachlan. But generative biology tools like RFdiffusion and ESM3 have quietly made text matching databases obsolete. We are not copying nature anymore. We are inventing entire evolutionary branches in a single afternoon.

Simon Carver

It, it, it really is wild. EvolutionaryScale ran ESM3 and asked it to build a novel green fluorescent protein. What came out had only fifty eight percent sequence similarity to anything found in nature. That is effectively bridging half a billion years of evolutionary distance in a single prompt. And then you have King and his team using genome language models like Evo to generate sixteen viable synthetic phages from scratch.

Lachlan Reed

Fifty eight percent? Blimey, so if the sequence looks forty two percent alien, the screening software at a place like Twist Bioscience just looks at it and goes, yep, looks harmless to me, pass it through?

Chris J. Murphy

Precisely. Legacy filters rely on text matching known viral sequences. If an AI outputs a three dimensional structural fold that acts like a lethal neurotoxin, but uses an amino acid sequence never seen on Earth, the legacy filter sees zero match and approves the print. The software is looking for known spelling, while the AI is writing a brand new language that delivers the exact same deadly function.

Simon Carver

And, and look at how fast the guardrails fall apart. Remember the MegaSyn study from twenty twenty two? Researchers took an AI model normally used to design therapeutic drugs to save lives, and they just flipped the reward function. Instead of rewarding non toxicity, they rewarded toxicity. In under six hours, running on a single standard laptop, that model generated forty thousand potentially lethal molecules.

Lachlan Reed

Forty thousand toxic molecules in six hours? On a normal laptop? That is, uh, that is properly terrifying, mate. Like leaving a trail bike in the shed and coming back to find it auto engineered into a missile. But surely the people holding the prompts are still responsible for what they hit print on?

Chris J. Murphy

That brings us to the core contradiction everyone is ignoring. Who actually owns the risk when the software designs something no human mind could ever predict?

Lachlan Reed

Well, look, I mean, I have to push back a bit here, CJ. To me, AI biological design software is just a sophisticated computer aided design tool. It is like AutoCAD for mechanical engineering. If an engineer uses AutoCAD to design a bridge, and that bridge collapses because the design was terrible, you do not sue Autodesk. You sue the engineer holding the stamp, or in this case, the technician holding the pipette in the lab.

Chris J. Murphy

The AutoCAD analogy breaks down the moment the software becomes agentic and non deterministic, Lachlan. When a civil engineer uses CAD, they understand every beam and load calculation. But when a biologist uses ESM3 or a model from David Baker's lab, they are handing the prompt to an engine that evaluates billions of structural folds beyond human cognitive capacity. The human is no longer designing. They are simply requesting an outcome and blindly trusting the three dimensional molecular behavior.

Simon Carver

Right, so when that synthetic enzyme or modified organism leaks out and causes ecological damage or harm, who gets sued? Is it the model developers at EvolutionaryScale? Is it the physical synthesis house that printed the DNA strand? Or is it the lab operator who just typed the prompt into a web screen? We have created a three way liability standoff where everyone points the finger at someone else.

Lachlan Reed

Well, but the model makers are trying to build safety net filters, right? Like, ESM3 open intentionally stripped out eukaryotic viral sequences during training so people couldn't generate dangerous human viruses.

Chris J. Murphy

That is the false security trap, Lachlan. Relying on training data filtering is like locking the front door while leaving all the windows open. Adversarial jailbreak frameworks like SafeProtein have already demonstrated that even with those eukaryotic viral filters in place, they can still achieve up to a seventy percent success rate in generating toxic protein variants.

Simon Carver

Seventy percent? So the digital filter gives corporate executives a warm fuzzy feeling, while anyone with basic prompt engineering skills can bypass it in an hour. We are relying on automated dashboards to save us from non deterministic biological tools, and it gives us total blind spots.

Chris J. Murphy

So what is the actual solution? It isn't better text matching filters, and it certainly isn't pretending human in the loop oversight works when humans cannot interpret the AI outputs. We have to pivot from sequence auditing to hardware level intent governance.

Lachlan Reed

Intent governance? What does that actually look like on a practical lab bench level, mate?

Chris J. Murphy

It means physical air gaps for biological design workstations, mandatory hardware level synthesis logging embedded directly into the physical DNA printers, and training human leaders to manage operational intent rather than staring at sequence compliance dashboards. You do not secure biological AI by reading the code. You secure it by governing the physical infrastructure and the human incentives behind the prompt.

Simon Carver

It really turns the whole biosecurity playbook on its head. We spent twenty years building databases to catch known bad sequences, only to realize the real risk isn't the sequences we know. It is the infinite landscape of functional biology we haven't even named yet.

Lachlan Reed

Yeah, fair call. Turns out trying to stop AI bio threats with text filters is like using a paper net to catch a ghost. Good chat, fellas.