I went on CBS19 TV news here in Charlottesville last night to talk about AI and biosecurity. I thought I was walking into a session that was going to be taped and played over a slow news day. I was pleasantly surprised (and a little thrown off) when I found out the segment was live on the evening news! Here’s the recording.
Biosecurity, and AI safety in general, is a really nuanced topic to try to get right in 240 seconds!
Here’s what I tried to get across.
The OpenAI / Hugging Face incident was a warning shot. It’s a cybersecurity incident that shows that capable AI systems can persist at hard problems, use tools, escape containment, and hack into systems people thought were secure. In biosecurity, AI can lower the barriers at important steps like finding technical information, planning experiments, troubleshooting failures, and operating tools. We need to test what systems actually do in realistic settings, in a safe way, to really understand the biosecurity risks posed by AI.
Be skeptical of doomsday countdowns. The anchor asked me about doomsday predictions so I advised treating these with skepticism. Some capabilities already here, but creating biothreats still requires intent, expertise, access to materials and equipment. And we here at the University of Virginia School of Data Science as well as many other really talented people and organizations are working tirelessly to make sure what just happened in cyber never happens in bio.
We should take the risk seriously without overstating it. Just because AI knows a lot about biology doesn’t mean some rogue AI agent can go out and create bioweapons. And, AI won’t turn someone into a biologist overnight: biology still requires materials, laboratory skills, and real-world infrastructure. We need to measure this to understand it. That’s one of the things we’re trying to understand here: what uplift does AI provide for doing biology if you don’t have a PhD in biology?
We need layers of protection. Defense in depth, as Steph Guerra and friends at RAND call it. Realistic testing before release, limits and monitoring for high-risk AI use, strong containment for autonomous agents, and safeguards at physical chokepoints like DNA synthesis and lab access. We also need independent evaluation and prompt incident reporting. We can’t rely on a single company to get this right for all of us.
Pace of AI progress vs AI governance. AI capabilities are improving fast. Governance needs to keep up. But we shouldn’t lock everything down with every possible restriction. We should prioritize measuring capabilities and how we mitigate threats: secure evaluation environments, independent testing, incident reporting, and screening at physical chokepoints. Without good measurement we can’t manage risk effectively.
It’s not either or. The anchor asked me: “Using AI to cure cancer would be good. Making bioweapons would be bad. What’s the balance?” Nobody wants to stop AI from helping with cancer, vaccines, or public health. We all want to make beneficial uses easier and dangerous uses harder or impossible. That means targeted safeguards, independent testing, and updating policy when the evidence changes.
Measurement matters. We can’t implement calibrated mitigations if we don’t know what we’re dealing with. Evals tell us a lot, but we’re also working to go further: when AI performs well on a benchmark, to what degree does that translate into a person being more capable at doing biology in the real world?