What do the people actually responsible for national security believe artificial intelligence can do? How are they weighing AI’s risks, including losing control of advanced AI, against the strategic imperative of deploying it? Last week, IST released the results of a first-of-its-kind survey aimed at understanding how national security practitioners and policymakers weigh these questions and more. The result is, we think, an unprecedented look at how the people who will actually make the calls view the risks, opportunities, and urgency of the AI transition.
We launched the findings at a live webinar, featuring a keynote from RAND’s Jim Mitre and a panel of the report’s authors and expert contributors. In this month’s newsletter, I compiled some of the most striking moments and interesting takeaways from the conversation into a Q&A format.
Q&A: Top takeaways, surprising conclusions, and next steps
Who did you actually survey, and why does that composition matter?
Philip Reiner, CEO, IST: “We surveyed a pretty broad and distinct group of 111 practitioners and experts. This included former under secretaries, retired general officers from the U.S. military, former senior directors from the National Security Council, and leaders spanning the whole of the U.S. executive and legislative branches. Crucially, 93% have engaged with AI policy at the organizational, national, and international levels. This makes it a very key group of interlocutors who actually are working with this every day. And so we tried to make sure that the survey was grounded with folks who actually engage with these technologies, and the ones that actually have to make decisions upon which these findings will weigh.
As you will hear today, these are not future decisions. These are decisions that really, in the national security context, are being made and need to be made now. What we found from their responses is, I think, an unprecedented look at how our national security leadership here in the United States views the risks, opportunities, and the urgency of this AI transition and what’s needed to do this securely, to do this safely, to ensure that we reap the benefits of these technologies, but also to have calculated understanding of whether to do so in the first place.”
Timing matters a lot for a survey like this. What was happening in the AI world while you were gathering responses?
Mariami Tkeshelashvili, Director for AI Security Policy, IST: “The timeline of the survey is very important, because it was launched [on April 30, 2026], right after Claude Mythos and Project Glasswing, and it ended right before the OpenAI and Hugging Face incident, when the autonomous AI agent broke inside third-party infrastructure. So that’s something to keep in mind while reading these findings: a lot has happened even since we closed the survey [on July 15, 2026].”
What makes this survey methodologically different from similar efforts you’ve seen in your work at the Forecasting Research Institute?
Bridget Williams, Associate Director of Research & Strategy, Forecasting Research Institute: “Two things about this survey stand out to me. The first is the sample. It’s a very experienced group with people across many different parts of the U.S. national security apparatus, including many people that are currently serving in the U.S. government. For us at the Forecasting Research Institute, this has been a group that’s been very difficult to reach in the past. So the fact that the IST team was able to reach this group, I think, is a real credit to them… And the other thing is the scale of the data that we have. There’s seventy questions across many domains, including many open-ended questions. So we have a very rich data set, and we’ve talked about a few of the findings today, but there are many more in the report than we were able to cover.”
What did the survey find about how close AGI-like capability actually is?
Mariami Tkeshelashvili: “Perhaps the single most important finding from our survey is that most respondents expect AGI-like systems within a few very short years, with artificial superintelligence following closely behind. Additionally, a sizable minority really fears that we are moving toward a future where it slips out of our hands, with potentially very harmful consequences. Respondents are very confident that AI will operate outside of human control within ten years. And when it comes to the risk of loss of control specifically, over half said that the risk of humans losing control over AI and causing a potential catastrophe is at or above the level they themselves deem acceptable.”
As the survey’s data lead, which result differed most from what you expected?
Ritika Verma, Associate for AI Security Policy, IST: “AI-enabled content analysis for threat automation was actually rated a critical risk by 10% of respondents, which was the lowest of all the options we asked about. This was interesting, because content analysis is something AI models are very good at, and it’s what we’re actually seeing happening in the landscape right now.”
What was the single most surprising finding for you in the biological and chemical weapons section?
Melissa Hopkins, Emerging Technology Advisor, Johns Hopkins Center for Health Security (CHS): “I think I was mostly surprised by the level of agreement on the risks from biological weapons development with AI. 70% said that AI already meaningfully increases biological weapons development risk, and 33% said that it will within two to three years – and that’s a lot of agreement. It’s certainly more agreement than most of the conversations I have with people who are building these biological AI models or who have wet-lab experience. And I was surprised that, at least it seems in the national security community, there’s more agreement on the risk than there seems to be in the broader biological community that doesn’t necessarily have national security expertise.”
If you had one to three concrete actions for policymakers coming out of this survey, what would they be?
Hamza Chaudhry, AI and National Security Lead, Future of Life Institute: “The first would be the distributional story around cyber. If you look at finding 40, a lot of folks are concerned about what happens with under-resourced defenders; they mentioned water, rural power, and municipal systems. I am really concerned that even in the best case, where we do deploy cyber tools effectively to large enterprise partners and the government, there will be many, many folks who don’t have those tools. I think it’s really key for us to set up a trusted access program and a deployment and diffusion mechanism that helps privilege those defenders.
The second is around the reality that the Secretary of the Treasury will be meeting his Chinese Foreign Ministry counterparts in 3 weeks in the run-up to the POTUS meeting with Xi Jinping. All of these questions need to come up in a track 1 at some point, and this can’t just be a domestic conversation — around CBRN, around cyber, around AI integration into high-stakes contexts like nuclear and military, and around biodefense, it’s very important that we start talking to the Chinese in detail.
And the third and final thing is to try to figure out what we can already do within the scope of the executive policy we have. Congress is slow, and these days it’s slower than usual. So, really interesting and entrepreneurial analyses of what we can do with existing executive authorities is something people should look into.”
