Security | Threat Detection | Cyberattacks | DevSecOps | Compliance

As AI agents spread across IT environments, organizations struggle to see what agents they have and what data those agents can access. For Cybersecurity Awareness Month, we wanted to share what Netwrix is doing to help IT and security teams improve agentic AI visibility and governance to keep their sensitive data secure. I recently sat down with Jeff Warren, our chief product officer, who has been working through this problem with customers for a while.

Researchers pulled real, leaked API keys out of three LLMs using only API output. Here is how AI agents carry secrets from developer laptops into model context, and how to stop them before the model sees them.

Risk in the agentic enterprise spans people, service accounts, and AI agents. All three act with legitimate credentials, so access controls alone can’t show whether their activity is safe. A behavioral risk framework applies one model to every actor: Verify expected behavior before access goes live, observe activity once it does, analyze behavior for risk and drift, and improve detection coverage over time.

An AI support agent picks up a customer case, signs in with valid credentials, works on it in only approved systems, and closes it. I would reasonably predict that if someone were to review this workflow, it would likely come back clean, with no policy rule violation. But if we were to take a closer look, we might find that the agent did more than just close the case.

This summer, we launched Identity Security for AI. That post described an upcoming wave of attacks and security challenges unique to Large Language Models (LLMs) and agentic workflows. As the traditional MITRE ATT&CK framework reached its limits, Google DeepMind published TRAIT&R, a taxonomy of rogue AI tactics and routines closely modeled after MITRE ATT&CK.

The platform determines what is known. AI helps determine what it means. Technology solution providers across virtually every industry are racing to add AI, in some fashion, to their offerings. Vulnerability management providers aren’t immune, either. They’re racing just as fast as the rest of them, sometimes faster. These vendors are embedding copilots, assistants, and agents powered by technologies like Anthropic Claude, OpenAI GPT, and Google Gemini.

Remote code execution (RCE) lets an attacker run commands on a system they don't own, usually through a flaw in a web application, CMS, VPN portal, or API. A successful exploit inherits the permissions of the vulnerable service, and many services run with far more access than they need. AI shortens the time from a disclosed flaw to a working exploit and lets attackers test thousands of exposed systems at once. A remote code execution flaw turns a vulnerable server into someone else's computer.