John Cumbers

AI Speeds Dangerous Bio Design; Economics Drive Safety Race‑to‑bottom

"What keeps me up at night is the fast pace of acceleration of open weight models in the ecosystem. They do not prioritize the same type of safety refusals." - Yunyun Wang, @OpenAI We just ran a biosecurity panel at @SynBioBeta today. Here's what you missed: AI is making it easier to design dangerous biology. The design-build-test cycle is compressing so fast the panel debated whether it's still a cycle at all. The screening tools mostly work. Kevin Flyangolts from @aclid said even heavily AI-redesigned sequences still conserve key functional residues, so current systems catch most of them... for now. But nobody wants to pay for screening. Jake Beal from @Raytheon said plainly that biosecurity is a pure cost center for every DNA provider. If you refuse to synthesize a flagged sequence, the customer goes to your competitor. Now, it's a race to the bottom. OpenAI deliberately downweighted biology training in their open weight models to reduce misuse risk. The problem is that the people building drug therapies and medical countermeasures need those same capabilities. The defenders lose too. Know-your-customer matters as much as computational screening. Scott Fay from @AnsaBio made this case. We've been talking about a passport system for legitimate users of risky sequences for years. Still doesn't exist. The Sequence Biosecurity Risk Consortium is trying to shift screening from "does this match a known pathogen" to "does this do something harmful to human cells." It's the right direction, but the speed needs to pick up. We love talking about what synthetic biology can build. We're less excited about what it takes to keep it safe. So who exactly is going to fund the defense side of this equation before something goes wrong? Something to think about. @DARPA @mkoeris #SynBioBeta2026 #SyntheticBiology #Biosecurity #Biotech