Sella Nevo spent years pushing the boundaries of artificial intelligence to save and improve human lives. His focus now is on developing policies to promote the safe and effective use of AI and biotechnology. His mandate at RAND: to help ensure the most pivotal technologies of our time benefit humanity rather than endanger it.
Nevo directs the RAND Center on AI, Security, and Technology (CAST). Its purpose is to provide evidence-based insights at the intersection of emerging technologies and national security—and to shape policies that protect global security while advancing the public good. That means identifying ways to foster beneficial uses of AI, including in cybersecurity, while preventing AI from creating the next bioweapon or launching the next cyberattack.
Nevo came to RAND from Google. He led a team there that used machine learning to tackle humanitarian and global development challenges, including forecasting floods. He's an expert in information security, especially as it applies to advanced technology. His work on protecting AI model weights, which encode everything an AI model has learned, set a standard for security now used in many frontier AI labs.
You've been at the forefront of studying both the promise and the danger of advanced AI. Let's start with the promise. What AI capabilities do you expect to make headlines a year from now?
I'll share some future capabilities that I'm excited about, and maybe they'll be here in a year. But no promises.
The most obvious one is on the coding side. We already see the models breaking out impressive capabilities there. One thing we don't have yet, but I think we might in a year, is world-quality software on command. The models will be able to build more, better, and much more personalized software.
We're also seeing an incredibly rapid pace of improvement in mathematics. If we could get 10 times, or 100 times, the pace of mathematical discovery, we could unlock a lot of solutions in security, material sciences, and other fields. It would far surpass any other era of incredible discovery.
Now the danger: What most concerns you when you look out toward that same one-year horizon?
The most imminent risk, even if it's not the worst-case scenario, is cyberattacks. We've seen rapid improvement in AI cyber capabilities, and the progress is not going to stop. Now, of course, this can also help defenders. But there are good reasons to think that, in the near term, these capabilities will be offense-dominant. It'll take time for defenders to catch up.
One specific moment that I think could happen in the next year is the diffusion of AI models capable of performing end-to-end cyber operations, with no skill required from a directing human—the AI just brings the whole package and compromises an entire business network by itself.
Just this week, we saw a frontier AI model autonomously hack an outside company to achieve its goals. These and even greater capabilities may eventually be in the hands of cybercriminals globally. At that point, all current bottlenecks for cybercrime and cyberattacks get removed, and we could see an unprecedented scale of cyber harm.
What do you see as the most important role for CAST?
CAST has been pivotal in recent years to U.S. policy, to lab security, to informing governments around the world. We're pretty uniquely placed. We're an organization that's deeply informed on the technical aspects of AI and biotechnology, for example. We also have credibility with the national security community. So we have a key role in bringing those communities together—tech, national security, and policy—in an area that is just rapidly rising in importance.
The most imminent risk, even if it's not the worst-case scenario, is cyberattacks.
You've also been working with METR on the Canary initiative, to evaluate the risks posed by frontier models. What have you learned?
When we got started with AI evaluations three years ago, AI models couldn't really do much. Now, they can work with cloud labs, end to end, with no humans involved, to design and synthesize a variety of potentially concerning products. Our evaluations show that they're getting closer to being able to synthesize pathogens on their own.
But perhaps more important than any specific result, we've been able to push the field of AI evaluations forward. We've helped inform industry and governments on what needs to be done to reliably and effectively evaluate open-source models, for example. We've also shown that incorporating agents into evaluations can reveal a lot of capabilities that were previously undetected.
What's a recent study that you thought was especially important?
We've provided both the U.S. government and frontier AI labs with recommendations for how to build highly secure data centers. These would be robust against attacks even by the most capable nation states. It's extremely difficult to protect massive digital systems against top nation states. They can spend billions of dollars and combine digital, physical, and insider attacks, undermining the security of even highly protected systems. Securing critical AI systems will be increasingly important as they become more capable. We've empowered key stakeholders to make these critical systems more secure than they've ever been.
What motivates you to do this work?
My primary motivation is actually pretty far from bio or cyber or even AI.
When I was 13, I took part in a program to support lonely Holocaust survivors. After my first afternoon with a survivor, seeing her appreciation, I realized: I actually did something that mattered today. I also realized that hadn't been true of any previous day in the first 13 years of my life.
Ever since then, I've tried to ask myself how I can do something that matters today. That has led me to a lot of different ways of trying to do important work. Eventually, it led me to try to tackle these tech and policy questions. That afternoon was really the event that most shaped the path that led me here.
