We Woke Up to AI and Cyber.
We're Still Asleep on AI and Bio.

When an AI broke into almost all of the NSA's classified systems in a single afternoon, people paid attention. Good. The lesson landed. But cyber is the risk we noticed, and it is not the worst one on the table. The one I cannot get anyone to look at is AI and biology, where the same jump in capability is quietly shortening the distance between an ordinary person and a pathogen.

A hacked network can be rebuilt. A released organism cannot be recalled.

One 2025 analysis estimated that AI has already raised the risk of a deliberate, engineered pandemic by roughly a factor of five. That number comes from two changes at once: AI makes the dangerous experiments faster, and it makes them better.

Faster, and smarter

Start with speed. In evaluations of Mythos, reported to be Anthropic's most powerful model, biology protocols that a trained team would expect to take about two months were finished in under two days. Sixteen hours of work, end to end. Steps that used to be the bottleneck stopped being one.

Now the quality. In one exercise, researchers split biologists into two kinds of teams. One kind were generalists with no special expertise in the problem in front of them. The other were specialists in exactly that problem. On a hard question, the generalist teams working with Mythos beat two-thirds of the specialist teams. The model can take someone without the training and hand them something close to expertise.

COVID is not the ceiling

A pandemic like COVID-19 is closer to the floor, not the worst case. There are engineered possibilities far more lethal, and the clearest one has a name: mirror life.

Every living thing on Earth is built from molecules that share the same handedness. Our proteins, our DNA, and the immune defenses that recognize an invader are all oriented one way. Mirror life would be built the opposite way, a chemical reflection of ordinary life. In theory it would function. In theory nothing in our bodies, and nothing in any animal or plant, would know how to fight it, because our defenses are shaped to grip the normal version and would find no handhold on the mirror one. Someone infected by a mirror bacterium might mount no immune response at all, and simply die.

This is not fringe speculation. In late 2024 a group of scientists, several of them Nobel laureates, published a warning in the journal Science urging that no one build mirror bacteria, on the grounds that there would be no defense against them. Several researchers who had been working toward the pieces of it stopped. The warning worked, which tells you the people closest to the science took the danger seriously. By most estimates at the time, actually synthesizing mirror life was at least a decade away.

We had roughly ten years of breathing room before mirror life became buildable. That estimate was made before you factor in AI.

The buffer was measured without AI

Today's frontier models, once safety training is in place, mostly refuse a bioweapon request, and what slips past still needs a competent, determined operator to be dangerous. Unsafeguarded, the uplift these models hand a motivated amateur is large, as much as fivefold in some capability evaluations. After safeguards, it gets much harder to use. I do not fully trust it.

My larger worry is AI accelerating biology as a whole, and building better general-purpose tools that can be repurposed. The same system that shortens a cancer project can shorten the road to a mirror organism, even if no one sets out to make a weapon. A ten-year buffer does not survive that kind of multiplier intact.

The response so far is thin. Chief executives and leading researchers have called for screening of DNA synthesis orders, so that a request for the genetic material of a dangerous organism gets flagged before it ships. America's 2025 AI Action Plan backed the idea. The pace of actually doing it is nowhere near the pace of the capability. Closed labs at least treat this seriously; open-weight models lag on biological uplift by maybe a year, and a year is delay, not safety.

Why this is really a superintelligence problem

Everything above assumes a human in the loop, using AI as a tool. It is already hard enough. It are racing to build systems that set their own goals and act on their own, and past a certain point, toward artificial superintelligence: a general mind more capable than us at everything, biology included, not the world the labs are building toward.

A system that compresses two months of lab work into sixteen hours, that turns a generalist into a specialist, that reads the route to a mirror organism the way it reads a paragraph, is exactly the kind of capability a superintelligence would have and would not need our permission to use. We already doubt we could control such a system. Biology is one of the most direct ways an uncontrolled one could reach into the physical world: quietly and at scale, against defenses that were never built for it.

That is why the Foundation's answer is prevention, not tuning. You cannot patch your way to safety against a mind that designs the pathogen and the workaround for your immune system in the same afternoon. The only protection that holds is to not build the uncontrollable system in the first place, and to put the same kind of binding, verified limits on the most dangerous AI that the world once put on biological weapons. Our plan is built around exactly that.

We woke up to AI and cyber because a headline forced us to. I would rather we wake up to AI and bio before the headline, not after it.

The strongest pushback

The fairest objection is that biosecurity already polices bad actors and that AI-bio is hype. Existing controls were built for a world without tireless model assistance. When tools compress skill and time, the old bar moves. Waiting for a release to prove the point is the wrong standard for pathogens.