AI is moving deeper into the enterprise. AI agents increasingly interact with applications, retrieve data, write code, and carry out tasks across connected systems.
That raises a question for security leaders. If an AI agent is compromised or begins acting unexpectedly, how far could it go?
In his recent commentary in Federal News Network, Louis Eichenbaum, Federal CTO at ColorTokens and former Chief Information Security Officer (CISO) of the U.S. Department of the Interior, argues that the risk lies in the security architecture supporting autonomous AI. Many organizations are expanding what AI can access without putting equally precise controls around that access.
Three Takeaways for Security Leaders
1. AI Access Is Growing Faster Than Security Controls
To do useful work, agents need access to data, software, and services. Some interact with cloud environments, application programming interfaces (APIs), source code repositories, and business systems. Yet enterprise networks were largely designed for people working through established applications and permissions.
The concern isn’t simply that an agent can act quickly. Broad permissions can allow one manipulated or compromised agent to interact with systems unrelated to its intended task. As adoption grows, these unnecessary access paths become harder to ignore.
Access the Guide: Enable AI Without Expanding the Blast Radius
2. Compromise Has to Be Part of the Plan
Traditional defenses remain important. But Eichenbaum argues that security cannot depend on every malicious action being stopped before it begins.
AI agents can be affected by prompt injection, model poisoning, software defects, and misconfigurations. They can also behave in ways their operators didn’t intend. Security teams therefore need to ask more than whether an agent is trustworthy today. They need to understand what remains exposed if that trust fails tomorrow.
That shifts the conversation from preventing every incident to limiting its consequences.
Access the Best Practices Guide: AI Threat Resilience in the Age of Mythos
3. Zero Trust Must Extend to AI Agents
Treating AI agents as identities is a practical starting point. Each agent should authenticate, receive permissions tied to a specific purpose, and be subject to ongoing evaluation. Access to one business function should not quietly become access to other, more sensitive systems.
Watch Video: Lou Eichenbaum on Zero Trust and AI agent access controls
Eichenbaum also points to software-based microsegmentation as a way to enforce granular communication policies around workloads and devices. By restricting which resources an agent can reach, organizations can reduce the chance that one compromised agent exposes unrelated systems.
The Bigger Question Is What Happens Next
The aim isn’t to slow AI adoption or pretend every new risk can be eliminated. It’s to ensure autonomy doesn’t come with unrestricted access.
Before expanding an agent’s responsibilities, leaders should know what it can access, what controls enforce those boundaries, and how much of the business remains reachable if it fails.
Are We Creating a Ticking AI Time Bomb? Read the full commentary on Federal News Network.
If you’re evaluating how to limit AI agent access and contain the impact of a compromise, contact us.
