Army Ends 100-Day Sprint, Plans Next Steps for AI

Fresh off 100 days of “sprints” aimed at eliminating roadblocks to AI deployment, the Army says it has a clearer picture of what it wants to accomplish by 2026, when it plans to become the first military service with a formal program designed specifically for AI.

Then there will be another 500 days of planning as the Army builds on the learnings from the first set of sprints and adds two new objectives to its roadmap: one initiative called “Break AI” and another called “Counter AI.”

Break AI’s direction is based on the assumption that the technology space will continue to shift toward what scientists call artificial general intelligence — models whose behavior can mimic human thought — rather than current technologies that tend to produce predictable outcomes using the same inputs.

“It’s about the concept of how do we actually test and evaluate AI,” Young Bang, principal deputy assistant secretary of the Army for acquisition, logistics and technology, told attendees at AFCEA’s TechNet conference in Augusta, Ga., on Wednesday. “As we move toward AGI, how do we actually test something where we don’t know the outcome or what the behavior is going to be? You can’t test it the way you test deterministic models, and we need industry help here.”

Meanwhile, counter-AI work is more focused on analyzing the capabilities the military will need to deploy to defend against adversaries’ use of AI.

“We want to make sure that our platforms, our algorithms, our capabilities are secure against attacks and threats, but it’s also about how do we counter what the adversary has,” said Jennifer Swanson, deputy assistant secretary of the Army for data, engineering, and software. “We know we’re not the only ones investing in this—a lot of the investment is in countries that are big adversarial threats to the United States. We’re not going to talk about all of that publicly because there has to be some operational security involved. But as we start learning and figuring out what we’re going to do, there will be things that we’ll share.”

New RFI to Defend AI Systems

Meanwhile, the Army is explicitly asking vendors to share what they know about securing and defending AI systems and technologies, one of the key issues the service has identified as necessary to address during previous 100-day sprints.

In a request for information the Army issued July 31, officials asked for feedback on its first version of the new AI Layered Defense Framework. Comments are due by the end of this month.

The framework — presented as an inverted pyramid — attempts to divide different types of AI models into three different risk categories: high, medium, and low, depending on how much work goes into securing them, as well as how they are used.

“We’ve identified about 30 different types of risks or attacks with about 60 different types of controls to help mitigate some of them, and we need smart people in industry and government to help us identify some of the gaps, scale and make improvements,” Bang said. “The idea is that as we go down the pyramid, we can establish more risk at the top. The lower you go, the more we want to accept less risk and have more control, and then let commanders and people make risk-based decisions. That helps us say, ‘Yes, we’ll put this in a weapon system,’ or, ‘We need more controls to address this type of risk.’”

The Linchpin project is expected to be fully funded in 2026.

These initiatives and more to come in the next 500 days will eventually feed into Project Linchpin, a new AI program the Army has been working on since 2023. The effort is partly a result of a program called Titan, an information fusion project the Army first demonstrated in the first iteration of Project Convergence four years ago.

“We had a pretty vague but specific requirement in our capability development documents that Titan had to use AI to reduce the cognitive load on the analyst, so we realized we needed AI,” said Col. Chris Anderson, project manager for intelligence systems and analytics in the Army’s program executive office for intelligence, electronic warfare and sensors. “As PM, I didn’t have a mechanism to really get it, and we had terabytes of exquisite data in the Army intelligence platform that was finally being consolidated into one place in the cloud for the first time ever. It was just screaming for some big language models to look at it.”

He added that it soon became clear that the same logic applied to the rest of PEO IEW&S’s portfolio.

“We saw a need for AI and machine learning, but we didn’t have a mechanism to do it. So a couple of my engineers and one of my product assistants came in and said, ‘Hey, we need to build an AI process.’ I said, ‘That’s a pretty good idea,’” Anderson said. “And then we took that up the flagpole to the Army, and what we have now is that it’s evolved to support the entire Army, not just my PM shop.”