AI Agents Create New Insider Threat Category, Security Expert Warns

Speed and autonomy change the threat landscape

AI agents operating autonomously inside enterprises represent a fundamentally different security challenge than traditional insider threats, according to cybersecurity veteran Shlomo Kramer. While human insider threats unfold over days or weeks with detectable patterns, an AI agent can execute thousands of autonomous actions before security teams notice anything amiss.

The distinction matters because most organizations are still building defenses designed for the old category of risk. As companies deploy AI agents that interact with users, other agents, data, and applications, controlling these interactions has become the primary security challenge enterprises face.

The recent Hugging Face incident demonstrated that an AI agent, when given a specific goal, can navigate around barriers intended to restrict it. Kramer argues the breach proves guardrails are necessary—the question is no longer if, but when they'll be implemented.

Why it matters

Every enterprise now operates AI agents with some degree of autonomy, and that number will only grow. The security implications extend beyond any single breach: organizations need visibility, governance, and real-time control systems built specifically for AI, not adapted from tools designed for previous technology generations.

Model builders shouldn't be security providers

Kramer, known in cybersecurity circles as the "godfather of cyber," emphasizes that this risk cannot be left solely to model providers. Whether frontier models or open-source systems, he doesn't expect model companies to provide cyber protection for the systems they build.

Cybersecurity has always been a specialized discipline requiring dedicated expertise. The team that builds a product is rarely best positioned to secure it—these are different disciplines with different mandates. That principle held true for enterprise software twenty years ago, and it applies to AI systems today.

Beyond nationalism and model wars

Framing the security challenge as open-source versus closed, or one country's models against another's, misses the actual issue, according to Kramer. The attack surface doesn't care about a model's origin. Every hour spent debating where a model was built is an hour not spent building controls that can prevent incidents regardless of their source.

National borders don't confine the challenges created by AI—they may actually exacerbate the technical, political, social, and economic obstacles that require collective solutions. Cybersecurity is the least of anyone's worries when compared to the broader challenges underpinning frontier AI development.

The case for global collaboration

Addressing AI risks requires global collaboration across model companies, security experts, governments, and enterprises. Each group holds a piece of the puzzle others don't: model companies understand their systems, security companies understand how attackers think and how enterprises get breached, and governments can set unified standards.

Kramer points to Nvidia's Open Secure AI Alliance as a step in the right direction, though just the beginning. International forums like the World Economic Forum offer platforms for diverse experts to tackle the complex governance, security, and policy challenges AI creates.

The question worth asking isn't which lab built the model or which country it came from. It's whether anyone is watching closely enough to catch what these agents are doing next.

These details were first reported in a Fortune commentary piece by Shlomo Kramer, founder of Cato Networks.