“We’re called the mad scientists of the Defense Department,” laughs Dr. Kathleen Fisher, director of the Information Innovation Office (I2O) at Defense Advanced Research Projects Agency (DARPA).

Founded by President Eisenhower in 1958 in response to the Soviet Union launching Sputnik 1 – the first artificial satellite sent into orbit around the Earth – DARPA’s singular mission has endured for over 60 years; to make pivotal investments in breakthrough technologies for national security.

Working with scientists, technologists and innovators both inside and outside the agency, over the decades DARPA has been responsible for developing many novel technologies that have gone on to become some of the most important technologies used in society.

Perhaps the most significant of these research projects started in the 1960s, when the Advanced Research Projects Agency (ARPA)– as the agency was then known – began work on ARPANET – a network for sharing digital resources among geographically separated computers. It laid the technical foundations for the internet as we know it today.

And DARPA continues to look towards the future by examining cutting edge technology and how it can be applied to national security.

“Our job is to create and prevent strategic surprise by looking over the horizon for disruptive technology in places where we can completely change the conversation and direction of what might be happening. It’s very high-risk, high reward work that other agencies wouldn’t take on,” Dr. Fisher explains.

“The range of the scope is anything to do with national security in the science and disruption space. Anything from how we build better aircraft to how we build better computers, or vaccines,” she adds.

While DARPA's mission is to push the boundaries of science and technology, it must be within the realm of reason, based on what’s potentially achievable with the technology of the day.

“Transporter tech from Star Trek is the example I often give; it’s something that would be really easy to make the case for why DARPA should do it, but we have a pretty good idea that it’s just not possible,” says Dr. Fisher.

Something DARPA is examining is AI-driven cybersecurity and how it can be used to enhance defenses against malicious cyber threat actors who wish to disrupt critical infrastructure and other systems vital to national security.

It’s this effort which is being led by Dr. Fisher, who is also a professor of Computer Science at Tufts University in Medford, Massachusetts, and a specialist in programming languages and their implementation. It’s her second role at DARPA, having previously served as a program manager on projects around high-assurance computing and machine learning from 2011 to 2014.

Typically, DARPA project managers only serve for short periods – typically three to five years – a system that’s designed to generate urgency to achieve success in less time than might be considered reasonable outside the Department of Defense. That means that at some point, Dr. Fisher will once again move on.

But for now, she’s focused on examining how AI can be used to improve cybersecurity of critical software – and setting some high targets for success.

“The grand vision is a world where software vulnerabilities are way harder to exploit; that it’s the exception rather than the norm,” Dr. Fisher explains.

“There’s this learned helplessness of ‘of course software is just helplessly vulnerable, and we just wait for the next bad guy to exploit it and that’s just the way it is’ – but we know how to build software to much higher standards, we are just choosing not to at this point.”

The thinking is that a lot of software development still relies on pushing out the product or update as quickly as possible, without appropriate thought given to testing and security – and that’s following a pattern which has been used in coding software for decades. DARPA wants to combat these legacy issues with the aid of AI.

“The motivation is the massive technical debt that we have accrued,” says Dr. Fisher.

“Part of the need to marry AI technology with cyber reasoning systems and traditional program analysis tools for finding vulnerabilities is just the sheer scale of the problem.”

“We can’t really rely on traditional techniques because we really need to find and fix vulnerabilities at scale; just finding the problems by themselves isn’t enough, we need to be able to fix them,” she adds.

Cybersecurity vulnerabilities in software affect every sector, but one which DARPA is thinking about is critical infrastructure.

The Cybersecurity & Infrastructure Security Agency (CISA) classes the systems and networks of 16 sectors as critical infrastructure – including communications, energy, food and agriculture, healthcare and water – on the basis that they’re "considered so vital to the United States that their incapacitation or destruction would have a debilitating effect on security, national economic security [or] national public health or safety."

However, a lot of the technology that powers vital critical infrastructure – utilities in particular –   runs on software that is vulnerable to cyber-attacks, either because patching the systems is too difficult, or because the software simply isn’t actively supported anymore.

“There’s lots of critical infrastructure areas and many of them are significantly underfunded. Being able to go to those sectors with solutions would be useful, especially the potential to not just find problems, but fix them,” says Dr. Fisher.

The desire to find an answer to this persistent challenge is the reason the agency launched the DARPA AI Cyber Challenge (AIxCC) in 2023.

The aim of the two-year competition is to bring together the best and brightest in AI and cybersecurity to design novel AI systems to secure critical code – with a total of $29.5 million in prizes awarded to the teams with the best systems.

The significant value of the prize money arguably demonstrates how important an issue the security of critical infrastructure is deemed to be. While critical infrastructure is vital to society, sometimes the budget – or motivation – isn’t there to ensure it is secure against malicious cyber threats.

But according to Dr. Fisher, not fixing the vulnerabilities will be a more expensive problem – and more damaging to society – in the long run.

"The cost of not fixing these vulnerabilities is very, very high. And it’s just going to get higher and higher," she argues.

DARPA hopes to provide a solution for this; and the agency believes harnessing AI could be the answer, which is why it’s running the DARPA AI Cyber Challenge (AIxCC).

For the AIxCC semi-final, teams - of which there were nearly 40 - were tasked with developing a Cyber Reasoning System, automatous software which can automatically identify and patch security vulnerabilities. These were tested against a set of Challenge Projects, designed and evaluated by experts and identical for each team.

Each of these Challenge Projects was based on a real-world, open-source project critical to infrastructure, national security or society and each contained vulnerabilities for the teams’ cyber reasoning systems to identify and secure.

All of this took place an environment built specifically for the contest at DEF CON 32; the immersive, interactive environment not only showcased live progress and scores in the competition, but it was also designed to look like a city and its utility providers– and to showcase the potential disruptive impact cyber-attacks could have on that environment to the thousands of DEF CON attendees who visited the contest.

“The stage was a mock city where we demonstrate the effects of a cyber-attack on critical infrastructure,” Dr. Fisher explains.

“The purpose was to create a visceral appreciation of the impact of cyber-attacks on critical infrastructure; it’s not just stealing people’s data. We wanted to drive home how important it is to pay down technical debt and produce an appreciation of what a cyber-attack could do to a city.”

Incidents like December 2015 cyber-attack against Ukrainian power companies – an attack later attributed to the Russian nation-state threat actors, which resulted in power outages in the freezing temperatures of winter – show the impact an attack on critical infrastructure can have on civilians.

The environment built for DEF CON in Las Vegas aimed to highlight the type of disruption a cyber-attack on critical infrastructure could cause closer to home.

The FBI’s warnings were echoed during the same hearing by CISA Director Jen Easterly.

“While the PRC is a sophisticated cyber adversary, many of its methods to break into our critical infrastructure are not. They don’t have to be. Why? Because we’ve made it easy for them. The truth is that, in many cases, the PRC is taking advantage of known product defects,” she said in prepared remarks.

“Unfortunately, the technology base underpinning much of our critical infrastructure is inherently insecure, because for decades software developers have been insulated from responsibility for defects in their products. This has led to misaligned incentives that prioritize features and speed to market over security, leaving our nation vulnerable to cyber invasion. That must stop,” she added.

For Dr. Fisher, the scale of the problem is clear, making it vital to secure the software behind critical infrastructure – potentially, if AIxCC is successful, with the aid of AI.

She points to the global Windows outage caused by a malfunction in a CrowdStrike software update as an example of what could be the scenario if a malicious hacking operation got inside critical infrastructure.

The update affected millions of systems around the world, causing flights to be cancelled, healthcare appointments to be rescheduled and payment systems to temporarily stop working, among other things. And it was caused by an error – a malicious attack might be much worse.

It’s because of the potential of threats like this that the DARPA AI Cyber Challenge is encouraging the use of AI-driven cybersecurity to help protect against malicious cyber threats.

Seven teams qualified from the semi-finals and they’ve each been awarded $2 million to help mature their technology ahead of the AIxCC Final Competition in August 2025.

The ultimate aim is to accelerate and facilitate the availability of software to find and patch vulnerabilities and defend what DARPA describes as “the software that enables modern life” – it’s a challenge which Dr. Fisher is excited to work on.

“I think it’s super exciting, because the size of the problem is enormous. AI offers a really promising possible solution in the future to find and fix vulnerabilities at speed and scale,” she says.