Research
Building a Defense-in-Depth Biosecurity Strategy for the AI Era
Aug 18, 2026
PodcastOctober 22, 2026
AI could help humans speed up cancer research. But it could also help bad actors create bioweapons. RAND's Steph Guerra breaks down a layered strategy to mitigate such risks while preserving legitimate scientific progress.
Deanna Lee
You're listening to Policy Minded, a podcast by RAND. I'm Deanna Lee. Earlier this year, scientists showed that trained AI models can design entirely new viruses. The experiment created viruses that infect bacteria, not humans, but this result is yet another powerful warning sign about the potential risks AI could pose. As AI technology progresses, it could open the door to unprecedented biological threats. The big question isn't just what's possible, but what can be done to protect against the worst-case scenarios while still allowing for legitimate scientific progress. Today, we're joined by Steph Guerra, lead author of a new RAND report that tackles that very question and lays out a strategy for biosecurity in the AI era. Steph, thanks for joining us.
Steph Guerra
Thanks for having me. I'm thrilled to be here to talk more about the work of our AI bio team at RAND Center on AI security and technology.
Deanna Lee
Excellent. We'll be getting into all of that. But first, tell us a little bit about yourself, your research, and what brought you to RAND.
Steph Guerra
So I'm a biologist by training. I got my Ph.D. in cancer research, and then soon thereafter transitioned to policy work. And actually about six months after I started a role in the federal government, the pandemic hit, and it was really an activating moment for me and an opportunity that allowed me to pivot my career to biosecurity and focus on preventing pandemics from ever happening again.
And so I've worked in biosecurity for the past several years, in a number of roles within the federal government, including the White House Office of Science and Technology Policy, and most recently at NIST, where I built out the federal government's capacity to evaluate AI models for national security vulnerabilities.
And so the reason I bring up both my biologist background and my work in federal policy is because I really appreciate both the promise and excitement that AI can bring to biological research, but also have a real grounded understanding of the risks through the work that I did at the White House and with the national security community. And RAND has been a really incredible place to be to continue this work. We have a great team of interdisciplinary experts focused on biosecurity, biotechnology—AI engineers, AI policy analysts—and all of them sit under one roof at RAND Center on AI Security and Technology, which allows us to leverage that expertise for the work that I'm gonna talk about more today.
Deanna Lee
You just kind of alluded to this, so I want to ask you, for our listeners who may not be familiar, what do we mean by the term dual-use biotechnology and why does it matter in the context of AI?
Steph Guerra
Absolutely. I think what we've been seeing over the past year especially is that AI is really reshaping biotechnology for good and for bad. And so what our greatest hope is, is that, AI systems and technologies like the chatbots, the AI agents, and even the bio tools like AlphaFold that produce DNA and RNA sequences instead of text. We hope that all of these different types of models will be able to help us learn new things about how our bodies work—how they interact with their environments—and pick up on the patterns and understanding of biology that we have not been able to understand as human beings.
Ideally, there's gonna be some sort of laws of nature that these AI systems are gonna help us to learn. And this can enable things like faster therapeutic discovery, new cures for diseases like cancer and Alzheimer's, and really overall reducing the resources and time that is needed to develop solutions to some of our greatest challenges.
Now, as you said, biology is what's referred to as a dual-use field. What that means in the context of AI-enabled biology is that the same capabilities that these different AI models have could be potentially used for good or for bad. So, for example, a model may have a really strong understanding and knowledge of biology and can do complex reasoning about biological systems. And that has incredible upside. That can support advanced research, diagnostics, and even education. At the same time, that same knowledge, understanding, and reasoning—if in the wrong hands—could enable harmful biological insights or novel design strategies.
Another example is we're increasingly seeing AI agents that are being developed, specifically called AI co-scientists that can assist in planning, execution, and discovery as it relates to laboratory tasks. And this will improve the efficiency, accuracy, and reproducibility of science, which is incredible. But in the wrong hands, it could also lower the barriers for conducting more dangerous experiments.
And so we really think about the risks at the AI-bio intersection in three primary buckets. First is what we refer to as "bioweapons uplift," so having capabilities that AI models can help a human by lowering technical, operational, or motivational barriers to acquiring bioweapons. The second is biosecurity circumvention, so helping an actor to bypass biosecurity controls to obtain materials that are part of bioweapons developments. And third is bioweapons novelty, which would be around using tools that can enable novel biological designs and create biological weapons that you can't find in nature. And those are some of the ways in which we think about the peril of AI-enabled biotechnology.
Deanna Lee
Now, in this report specifically, you focused on the worst-case scenarios: events at least 10 times the scale of the COVID-19 pandemic, which I think for a lot of us probably feels a little bit hard to imagine. You just gave us a little bit of a description about what these three buckets are, but can you paint a little of a picture about what a worst case scenario might look like that is enabled by AI?
Steph Guerra
Absolutely. So, for those of us that work in public health and biosecurity, these types of worst case scenarios—something that is at the scale of 10 times the COVID-19 pandemic—are actually not too hard to imagine because we do have historical precedence for these types of events. So the COVID-19 pandemic killed a confirmed about 7 million people worldwide. When we look back just to 1918 and the influenza pandemic then, we had an aggressive H1N1 influenza virus that infected roughly one-third of the global population and killed an estimated 50 to 100 million people in less than two years.
So that in and of itself was a natural pandemic that emerged that was able to kill the population at a scale of, you know, 10 times what COVID-19 enabled. Going further back, of course, we have examples like the Black Death, which was the deadliest pandemic in human history caused by the plague, and that killed an estimated 50 to 200 million people across Europe, Asia, and North Africa. Other examples include smallpox, which killed an estimated 300 to 500 million people worldwide during the 20th century alone.
Now, of course, all of these are historical examples—smallpox, in particular, was eradicated by the vaccine. But what all these devastating events have in common is that they're caused by transmissible diseases that have the right level of pathogenicity, so they're able to kill people, but not every person they infect. They also have a high level of transmissibility, so they're spread rapidly throughout the population. And maybe most importantly, they have novelty to the human immune system. So these diseases, when they emerge, oftentimes, human beings that are exposed to these diseases at the time have no natural immunity. And because of that, they're able to spread and kill quite easily.
And so, what we're concerned about now with AI-enabled biotechnology is that it could make the possibility of future pandemics not just come from natural emergences but also come from other sources like accidental laboratory leaks or engineered human-caused pandemic. And so what we want to better understand is how AI can make it more feasible, plausible, for humans to design something novel—that novelty piece that I talked about—and also just generally get their hands on making these types of viruses de novo. So with the advancement of AI and with the enhancement of biotechnologies and their convergence, what we're concerned about is that there will be more people that are going to be able to upskill to the level of expertise needed to be able to create these viruses in the first place and cause a pandemic. And so it's not just nature we have to worry about now, but also nefarious actors who could possibly cause this level of devastation.
Deanna Lee
Okay, so we have a better idea now of the risk that we're dealing with when it comes to AI and biotechnology. And your research proposes a layered defense strategy to counter those risks. Can you give us just a quick summary of what that means and why this layered approach is sort of the essential way to tackle this risk?
Steph Guerra
Yeah, absolutely. In biodefense, we oftentimes are focused on developing what we refer to as a defense-in-depth strategy. And the reason why we focus on having many layers to this strategy is no single intervention is going to be bulletproof. So, this strategy lays out nine different mitigations that, when layered on top of each other, provide meaningful friction across the entire risk pathway in order to detect, disrupt, or even deter a threat actor from pursuing a biological attack in the first place.
This multi-layered approach is important because each layer alone can be potentially circumvented. So... one of the examples in the report is a very well-known layer of biosecurity, which is called gene synthesis screening. And what gene synthesis screening is and means is that scientists that are doing legitimate research often have to order pieces of DNA or RNA in order to conduct their legitimate research. And to do that, they order those fragments from gene synthesis providers. Oftentimes, the vast majority of orders that go into those companies are going to be not harmful constructs. However, you know, it's possible to order fragments of more harmful construct like fragments of the influenza genome that can then be cut and paste together.
And so, gene synthesis screening is an approach where these companies will screen the orders to ensure that they're not provisioning someone with fragments or parts of pandemic viruses, and know-your-customer screening, which ensures that they provide the materials solely to legitimate customers. Now, the reason I bring this one up is it's a great example of the fact that this is a very important and critically necessary mitigation, but it won't be sufficient. Because there are some actors that don't rely on ordering fragments of DNA or RNA from commercial channels. They may be situated within a state biological weapons program and have their own in-house capability to do that. Or, there are ways to circumvent this screening by potentially placing orders at many different companies so that you don't actually trigger a signal at one company that makes them not want to provide you that piece of DNA or RNA.
I think the same goes for other mitigations that we recommend, and the strategy is that there's always going to be a way for a motivated actor to potentially circumvent that mitigation. But if we put up enough barriers along the entire risk pathway, it will lengthen the amount of time, raise signals that could lead to disruption, or just make the actor realize that it's not actually worth it to pursue this path in the first place. So, that's sort of the reason why a multi-layered approach is important in this, because it's so hard to lock down the materials and information needed to develop a biological weapon, because it's such a dual-use science, like I described earlier.
Deanna Lee
Right, and I think the example you just provided, which was excellent, really demonstrates some of the complexities of addressing risk here. You mentioned that there's nine mitigations in the strategy that's outlined in your report. We're not gonna walk through all of them, but there are some kind of big themes that emerge, and you just hinted at these a little bit. One of those is that different types of threats and different types of actors require different defenses. So an intervention that will work well against a lone individual isn't going to be effective against a nation state that's pursuing an AI-enabled bioweapon and vice versa, of course. Can you talk a little bit about the need to have mitigations that sort of focus on the variety of different threat actors that could be posing risks?
Steph Guerra
That's correct. And, to go back to my last example—access controls like gene synthesis screening, which provides a barrier or blocker to a nefarious actor getting the physical materials for a biological weapon—is one approach that will only work on an actor that doesn't have access to their own native capability to access DNA or RNA. Access control is also something that won't work on a legitimate scientist that somehow gets radicalized. So like a great example here is from the Amerithrax attacks from early 2000s, where the alleged perpetrator of that attack was someone who had legitimate access to anthrax through his research and work. And so it's an insider threat. And so, insider threats likely would be more challenging to disrupt through access controls because they have the access already.
There's other types of actors like a sophisticated state biological weapons program—those actors again, won't be affected by access controls, but may be affected, by deterrence layers. So it's possible that a nation-state is not going to want to use a biological weapon if they think that there will be consequences for that, that they will be attributed, that they'll be caught, that there will be challenges for them geopolitically. And so for that reason, we have other layers in the strategy that are more focused on deterrence. So generally what we find is that all actors can either be disrupted in some way or deterred in some way, but oftentimes it's hard to do both to the same actor because of their access to resources, expertise, or their just motivation differences.
Deanna Lee
Okay, and I want to give you another example that I think might come to mind readily for people when they think about AI and biosecurity. So, let's say a user asks an AI chatbot or an AI model how to maximize harm with a bioweapon. And I think the most natural defense we all might think about is that the model simply wouldn't engage, right? But your report argues that this idea of blocking isn't really sufficient, right—that monitoring what users are doing is also really important to add an effective level of defense. Can you talk about the need for both monitoring and blocking and how they kind of complement each other?
Steph Guerra
Yeah, absolutely. So, we argue that an effective strategy requires a complementary detection architecture that can identify and disrupt misuse patterns in both the digital and the physical domains before they culminate in harm. So this helps to shift this paradigm from this sort of just denial strategy to integrating monitoring and early response.
So there's several advantages that misuse monitoring can have over blocking access to biological capabilities from these AI models. It can allow better detection, deterrence, and risk reduction than blocking alone. So, for example, when a query is blocked, the threat actor is alerted that their approach has triggered a safeguard, which can give that actor the chance to adjust their behavior, or even change their behavior and go to a different environment where monitoring is not possible, such as potentially an open-weight model that doesn't have these safeguards in place. And this reduces overall visibility into that malicious activity.
And so a monitoring-based approach, by contrast, really inverts this because the threat actor is not alerted, they don't change their behavior, and this could provide us more useful information about future misuse. So it really lets organizations set their detection threshold more inclusively than would be acceptable for blocking alone. And it also does not result in an immediate user-facing action. So, it just is a complementary approach. I think this approach, when deployed appropriately, can be a fundamentally stronger mechanism for catching genuine misuse and reducing overall risks.
But there's a number of things that we have to get right about this. Some of this includes actually understanding the distinction between when someone is using a model for good biology versus bad biology. That's this dual use dilemma that's been a challenge in biosecurity for decades. What I find hopeful is that now we have a lot of information about how legitimate scientists are working on virology, immunology, and microbiology, often through trusted access programs that these AI labs are implementing. And so we can start to develop baseline models of what normal biological information seeking is, and then try to see where anomalous patterns can be identified.
And so this is, sort of like, is still an area of active investigation by the AI labs themselves. And it really requires collaboration between the AI developers, the AI deployers, the users and the scientists and biosecurity specialists to start to understand and distinguish between legitimate use and also potential misuse. I think the most important thing, though, to talk about when it comes to this as a potential strategy for adoption is how we can do this in a way that maintains privacy. And so, creating privacy-preserving tools that don't expose individual data to human review unless it actually goes above a threshold is something that's really a prerequisite for this work to proceed in a manner that can benefit both biosecurity and also legitimate research.
Deanna Lee
Excellent. I'm really glad that that you brought up privacy. That's obviously a major consideration for listeners and surely for policymakers as well. So we've talked about a few of the mitigations. Are there any that might surprise people or anything that you feel isn't getting enough attention, especially given all of the headlines about AI risk right now?
Steph Guerra
So, gene synthesis screening is perhaps the most well-known mitigation that we describe in this strategy. This makes sense to me because it really acts at a very important physical chokepoint where someone's ideas and designs from the computer come off of the computer screen and actually manifest in the real world. What we add to this conversation in our report is talking about how gene synthesis providers are not the only source of potential biological weapon precursors. So, we discuss other emerging suppliers within the biological supply chain, including contract research organizations, automated cloud laboratories, even plasmid repositories that exist in academic environments. All of these are also potential sources of the—they're referred to as plasmids or vectors that contain the DNA or RNA fragments that make up a virus. And so that is an important part of expanding this physical chokepoint across more areas.
I think the perhaps more surprising part of our report is our proposal to also add digital chokepoints. So, trying to move some of the layers of biodefense further upstream, so not just at the point where you've decided to start building something, but also to ensure that someone can't actually get the blueprints for the novel pathogen in the first place. And so these additional layers include protecting sensitive biological data, so really well curated laboratory validated data around pandemic influenza. And then the second piece of the digital chokepoint, or sort of managing information access, is around managed-access platforms—where these tools and data sets can live and be accessed by the legitimate researchers but not openly publicly available to be potentially misused. So I think that sort of digital layer of defense is something new that we bring to the discussion. And we're looking to find what are the other layers besides gene synthesis screening that we can build technical grounding to through our analysis and potentially widespread support across the biotechnology ecosystem.
Deanna Lee
Really interesting. And you've touched on this a couple of times now, but I do want to ask you explicitly about the tension that's in the report between protecting dangerous information from getting into the wrong hands and preserving open science. And I'm sure this is something you're very focused on as a cancer researcher. And it's also something that's kind of been defining the broader debate about AI recently. How do you think about that trade-off in this research?
Steph Guerra
Yes. So our team thinks about this a lot. And the majority of our research and analysis on AI safeguards, we refer to as innovation-compatible safeguards—trying to find the right approaches and the right level of detail on what exact types of data, what types of tools have the highest potential misuse risk. And then finding ways to put those tools and data into the hands of people who do need them to develop medical countermeasures, but being realistic about whether or not they should be openly and widely available to anyone on the internet.
And so there is a trade-off here, but there are ways and approaches that we can build biosecurity by design into how we build and use these tools and datasets. So, our RAND team last year put out a global risk index that looks at AI-enabled biological tools and assesses them for their misuse potential. And I think—spoiler alert there—it's not every tool out there that has the capability to design or develop a novel pathogen. It's a small subset of tools that have that misuse potential. And we're looking to better define this for biological data as well.
And so, we really care about preserving open science. We've been excited so far by our initial work that seems to showcase that it's really a small subset of the tools and datasets that would potentially require more stringent access controls. And we've also been encouraged by what we've been seeing across the community with Frontier AI Labs launching managed-access programs for their biggest and most powerful biological partnerships with organizations like CEPI—who build medical countermeasures to prevent and respond to future pandemics. They've just launched what they're calling the Pandemic Preparedness Engine, which is a managed-access platform for the tools and data and technologies that their researchers are building. So this is something that can be built in from the beginning of any large scientific effort to advance AI-enabled biotech.
Deanna Lee
Alright, so we've talked about the potential misuse and the risks. We've talked the layered defense strategy. Let's talk about implementation. What's the first step to putting this strategy in place?
Steph Guerra
So this strategy has nine different mitigations, and I'd say each of these mitigations has their own unique next steps. So, some of them fall into the category of being more nascent, maybe more theoretical, where we really need more research and investment in order to determine is this a worthwhile safeguard. So, an example here would be we propose the need for tamper-resistant safeguards for both closed and open-weight models, which means those sort of refusals and blocking that you talked about earlier, having those so that it's not possible for them to be jailbroken or removed. You know, there's been some early research into this as an approach, but not nearly as much investment.
The second category of mitigations fall into what we refer to in the report as technically feasible, but still early development. So there's been a lot of technical work on this and more is needed in order to start to implement this worldwide. Next is established, but have some implementation gaps that remain. And, you know, a great example here would be gene synthesis screening. And then lastly, we have technically tractable but institutionally complex. So these are things that—we don't really need a new novel technical breakthrough—it's not rocket science to get this done. But to actually build a new institution or make some of these ideas come to life in the real world will require really strong proactive leadership.
And so I think there's a lot of important considerations out there for policymakers. The first thing I'd say is that many of the approaches that we've described here for the intersection of AI and biology are also relevant to other domains. So these are often cross-domain issues that require cross-domain solutions. So all of the mitigations that we describe in this strategy are going to require credible and sustained U.S. leadership to make happen. U.S. leads the world in AI development, and we have to start making some decisions around deployment of these AI systems to enable all of the beneficial things that they're poised to achieve while also reducing the risk. And partnership with industry is also going to be vital. They're the ones that oftentimes have the most advanced technical development of safeguards, of managed-access platforms, of real-time monitoring. And so, providing standards across what is considered an acceptable, responsible way to deploy AI is particularly important for governments to take the lead on.
Deanna Lee
And what would you say are the biggest obstacles to taking action or to implementing these mitigations and maybe some ways to overcome them?
Steph Guerra
I think one of the biggest obstacles to implementing these mitigations is the fact that understanding the misuse risks and being confident that an AI model can be misused to cause a bioweapon attack is a really hard thing to do. So, about half of my team at RAND focuses on evaluating and understanding these capabilities and risks, and the other half is focused on development of mitigations, such as those spelled out in the strategy.
The reality is, we likely will never have full confidence that an AI model can cause a biological weapon attack, because it's really hard to measure that. It's not as easy as cyber capabilities where you can have a cyber sandbox of some kind and actually test things, like how good the models are at hacking or performing AI agent swarms and things of that nature. In biology, we can't do that same thing. It's not responsible to actually see whether someone can develop a biological weapon with AI. And so what we do is we actually take pieces of that risk pathway and perform evaluations often on proxies or non-hazardous materials in order to better understand the risk.
And if we wait too long until these capabilities are demonstrably effective, it will be too late. Everything that we've described in this report will likely take years to implement properly. And so, there's not really time to wait. So I think making these investments in times of deep uncertainty and sort of some lack of confidence in the capabilities that are being observed is one of the major challenges and the biggest obstacle to actually implementing them in the first place.
Deanna Lee
Absolutely. The report also emphasizes the importance of international coordination, especially between the U.S. and China—the two big players in the AI race, as it were. How important is a dialogue between Washington and Beijing when it comes to putting this strategy into practice? It's not something the U.S. can kind of go at alone, right?
Steph Guerra
That's right. Cooperation between the U.S. and China on biosecurity measures like the ones outlined in this strategy is vitally important. Even if the U.S. was able to implement all of these measures perfectly, there would still be a large gap in our ability to disrupt, detect, or track threat actors or future outbreaks if we don't have cooperation with China and others across the world. We emphasize China in particular because they are our peer nation in both the fields of AI as well as the field of biotechnology. So if the U.S. was to implement mandatory widespread gene synthesis screening, but China or even other nations around the world did not, that would just leave open, in our global economy, other places people could go to get these precursor materials. So, this cooperation with China and with other nations around the world is vitally important.
Deanna Lee
Before we go, I wanna ask you, what's one thing that you want listeners to take away from this conversation? Because we've all been a little overwhelmed by talk about AI risk lately. So what's your key takeaway?
Steph Guerra
My key takeaway for this report is that the time to act is really now. These mitigations will require cooperation with industry, cooperation with China, investment in technical solutions, industry partnerships—all of these things are non-trivial to implement and likely will take quite a while. And many of the mitigations in this strategy are those that we have been calling for in the biosecurity community even before AI entered the zeitgeist. These are, you know, well-trodden, important approaches that we have a bevy of experts out there ready and willing to help with implementation and make them happen.
I think what gives me hope about our ability to prevent future AI-enabled biological incidents is the amount of people that have entered this field and really see this work as a mission. A lot of people out there, myself included, entered the space of biodefense after the COVID-19 pandemic, seeing the suffering and wanting to ensure something like that will never happen again. And it gives me hope that there's a really impressive and devoted workforce that wants to make many of the things in this strategy a reality.
Deanna Lee
Well, I think that's an excellent note to end on. Steph, thank you so much for your time and for your insights today.
Steph Guerra
Thank you.
Deanna Lee
And thanks to our listeners. You can find a link to the report we discussed today at rand.org/podcast. Today's episode was recorded and edited by Emily Ashenfelter. It was produced by me, Deanna Lee. RAND's director of digital communications is Pete Wilmoth.
RAND is a nonprofit institution that helps improve policy and decisionmaking through research and analysis.