Artificial intelligence (AI) is enabling faster, more widespread cyber-attacks that can overwhelm defensive systems. Maintaining the status quo is not enough, as preexisting vulnerabilities in legacy systems and gaps in operational networks are quickly identified and attacked by adversaries who have faster and more adaptable tools at their disposal.
Breaking Defense discussed how cyber defense is evolving to protect critical systems with Leidos’ Paul Welch, Senior Vice President, Business Area Lead, Defense Agencies, and Josh Salmanson, Vice President, Defensive Cyber Practice.
Breaking Defense: Describe the threat scenarios necessitating the need for defensive cyber techniques, especially those that are more proactive than reactive.
Welch: Let’s start with the 800-pound gorilla: AI. The use of AI by threat actors – both criminal and state sponsored – is increasing the risk across cyberspace in general. Specifically for the Department of War, as data becomes increasingly important to the department’s operations, the importance of that data availability and integrity also increases, in some cases exponentially. When that data is not available or perhaps cannot be trusted, the risk to operations rises dramatically.
Couple this increasing reliance and demand for data with the complexity of modern infrastructure and combat systems, along with the growing threat of AI-enabled adversaries exploiting vulnerabilities with those systems, and the need for a more proactive approach to cyber defense that is focused on that mission integrity becomes apparent.
Salmanson: It really starts at the beginning of the life cycle. We need to build systems that are more resilient by design. Too often, critical systems remain fragile because mission demands are continuous, leaving little time to address the technical debt that accumulates over time.
That fragility creates downstream risk. When a new vulnerability emerges, organizations may rush to patch systems at accelerated tempos, sometimes without fully testing the patch. In complex environments, those patches can unintentionally disrupt other mission-critical functions, creating operational instability.
The broader intention is to move much faster while also building more secure and resilient architectures from the start. Modern development, automation and security engineering practices are intended to help organization keep pace with increasingly AI-enabled threats. The things that we’re seeing right now, especially with the emergence of Mythos-class models and some other capabilities, suggest that operational tempos for defensive cyber activities may need to increase substantially.
What agencies and industry are seeing is these advanced models can identify vulnerabilities and exploit paths at speeds and scales that were previously difficult to achieve. In many cases, they are rapidly surfacing weaknesses that organizations did not realize existed. As a result, defenders will likely be forced to patch things faster. That’s going to put a tremendous strain on the operational and engineering teams. The challenge now is how to figure out how to respond to that increased burden without introducing instability into already complex environments.
How can we impose costs on adversaries and their cyber offense activities so that they’re slowed down?
Salmanson: In this context there are a lot of different ways defensive cyber operations can impose costs. Primarily, deception is a fantastic way to go. It’s part of every other military domain. The U.S. military has used deception in warfare forever and yet we haven’t really applied it in cyber.
It’s critical, because it’s the one thing that slows down the adversary in the absence of any real perimeter anymore. Identity is the new perimeter, and there are many, many more identities than people that use the systems now. Organizations must manage system accounts, automated services, AI activities, API integrations and cross-cloud connections. All those identities expand the attack surface and blur the boundaries of what is truly “inside” the environment.
As a result, it has become significantly harder for agencies and enterprises to distinguish legitimate activity from adversarial behavior. In many cases, the most effective way to identify an intruder is to introduce carefully designed deceptive elements into the environment – decoy credentials, false data, synthetic services or other lures that appear operationally valuable to an attacker.
If you’re using deception, they have to get in for you to deceive them. Is that not a bit of a failure?
Welch: They have to gain access to the part of the enterprise you want them to enter. There are many mechanisms and tradecraft techniques adversaries use to do that. You can channel and direct where an adversary goes by making a target appear attractive, interesting, and more accessible to them.
But where they’re actually going or potentially going is ultimately somewhere that is not useful to them. You’re imposing cost because the information they’re accessing is not real. They’re wasting their time and they’re exposing their own tactics, techniques and procedures because you’re observing them the entire time.
Salmanson: A lot of times when an adversary gets caught in a deceptive lure, they’re no longer on the target network that they were looking for anyway. They’ve been moved off into a very safe space where their tools, tactics and procedures can be observed. The adversaries think they’re still in the real environment, yet they’re in a hall of mirrors that’s totally under our control. If they attempt to get back into the real environment, we know how they got in the first time, because we can go back and track all of that forensically. Cyber deception does help and it does impose a tremendous amount of cost when it’s done right.
We talk about the enterprise as one monolithic thing, but obviously it’s not. What are the challenges of cyber defense when there are multiple enterprise networks?
Welch: Multiple infrastructures and multiple technologies within an infrastructure that are all interrelated and communicating [all have] seams and present vulnerabilities in those seams. You have multiple hosts and compute environments that also introduce vulnerabilities because of the seams and multiple applications that have multiple interactions.
More importantly, or equally as important, you have 40-plus defended areas under 40-plus purviews in terms of defensive cybersecurity providers across the entire department. The ability to gain a singular appreciation and a singular picture of situational awareness over all of the activity within the domain is virtually impossible.
[It’s important to] understand what is occurring within the Air Force portion of the enterprise versus the Navy portion of the enterprise when there may be activity that’s related across those domains that defenders need to be able to correlate. Then defenders must have a coordinated ability to respond to that activity. Even if just reacting to the adversary, coordinating security actions is important, when defensive operations shift to a more proactive approach, that coordination becomes critical. This need to enhance the ability to coordinate across the department is a priority of leadership and drives much of the innovation we bring to cyber operations.
You mentioned seams. What’s a seam in this world that we’re talking about right now? Why are they vulnerable?
Welch: Any time you have a warfighting force operating next to another warfighting force there is a potential gap in those forces’ operational areas. If you know where there is a seam between those two forces, then you have an operational vulnerability. How is it that you make sure that you’re coordinating the two? Closing that gap is all about command and control, having shared situational awareness of the activities and the lines of operation, and having a shared understanding of commanders’ intent and execution.
In what other ways is AI changing the cyber equation on both sides?
Salmanson: The biggest problem that defenders have today is that the volume and velocity and variety of attacks have changed dramatically and it gets worse every year. It’s an absolute hockey stick right now from the exponential explosion of attacks. It’s hard for systems to keep up.
The challenge is this: the environments that almost all of us have to defend are based on older architectures that have been sustained for two, three generations now. The industry keeps making incremental progress, which isn’t enough to face an exponential attack cadence. Organizations must change their approach to meet the level of threat. We are often making decisions based solely on cost, not on operational need.
My belief is that organizations and industries have to change the core hardware that runs enterprise software so defenders can keep up with the type of attacks that they’re seeing today and be able to respond much quicker. When the adversaries are driving Ferraris and the defenders are still driving a Ford Fiesta from 1982, it is clear the defenders are going to have some problems.
The industry is starting to make some very subtle changes, but unless you’re on the biggest platforms, you’re not getting the benefit of the exponentially faster hardware today. The difference between a CPU and a GPU is about a thousand times faster. If you’re going to fight, you might as well fight with the same advantage that your adversary has.