JD Durkin: Joining us now, Yuriy Bulygin is the founder and CEO of Eclypsium, whose new, recent research may expose a bit of a blind spot in the current ongoing AI boom. Yuriy, grateful for your time on this Friday. What a time to wrap up this day on Wall Street, this week on Wall Street, to be having this conversation. Let me start with this: AI investment expected to top $1 trillion this year. Do you think companies and investors might be overlooking a bit of a security risk in the infrastructure that powers the AI boom?
Yuriy Bulygin: JD, pleasure to be here. Absolutely. In fact, much of the investment and discussion in the cybersecurity industry today related to AI is happening in the upper, logical AI layers: how to protect AI applications, workloads and the agents, autonomous agents, and how to prevent prompt injection attacks and data from being exposed, which are all legitimate and important topics of conversation and protection.
However, all of the AI workloads and applications and data and models are running on massive, massive infrastructure. And like I mentioned, there’s a north of $1 trillion of investment happening in building that infrastructure. And all of the future infrastructure is going to be AI infrastructure. So the conversation also needs to include how do we secure that infrastructure? Because there is no trusted or secure AI without trusted infrastructure that it runs on. And this is what our team has been investigating, and also highlighting some of the risks that are inherent to that infrastructure that we need to take care of.
Because if attackers can compromise that infrastructure, they can steal the data. They can modify the outputs of those models. They can modify the behavior of those AI agents that we are relying on in our personal lives, in the enterprise, in protecting our organizations, and even in national security. So it is important to shift the conversation, or expand the conversation and investments, into how do we build the infrastructure for AI in a secure way.
JD Durkin: So your new research shows that attackers can actually change the way that an AI model behaves in real time without even touching the model itself. How would that work? And from where you’re situated, why is that so concerning?
Yuriy Bulygin: Absolutely. This is what we have demonstrated. And imagine, we all rely on AI in our personal lives, maybe when we’re asking our chatbots to help us with some kind of health symptoms; in our organizations, maybe when we’re developing AI-based applications and workflows, and we need to make sure that those are developed in a secure way; or maybe in national security, when we’re doing an assessment or some sort of potential military operation or some other actions critical to national security.
Now, what we’ve been able to demonstrate is that sophisticated adversaries and determined adversaries, maybe nation-state adversaries, maybe other threat actors, are able to take over some of the infrastructure that the AI models and AI agents run on. And by doing so, they now control the operational environment for all of those models and agents, and they’re able to control the outcomes of those models and agents, including modifying the outputs that those agents gave us, including modifying the behavior of those agents. For example, if you’re developing an application, the compromised infrastructure can influence the model and agents to build some kind of backdoors or vulnerabilities or other bad things into the code that is generated by those agents.
Or if you’re asking about certain symptoms, the attacker is able to modify the model behavior in real time to give a wrong, potentially dangerous output.
JD Durkin: Wow. Yuriy, it’s a fascinating time to be having really important conversations like this. Unfortunately, that’s all the time we have today. But please come back and join the broadcast again, because this is certainly a topic we are going to be talking more and more about as the AI boom continues to be built out. Yuriy Bulygin, founder and CEO of Eclypsium. I’m grateful for your time.