/
Segmentation

New Research Shows Why Segmentation Is the Answer to AI Agent Security

In the UK, nobody gets the car keys on their 17th birthday and a cheerful wave toward the motorway.

New drivers start with red L plates on the bumper. They can’t drive alone, and the person in the passenger seat has to be at least 21 years old and have held a full license for three years. Only after plenty of supervised practice and passing both a theory test and a practical test does anyone get to drive on their own.

The UK built that system because driving is consequential. One bad decision at speed is hard to take back.

Now think about what many organizations have done with AI agents over the past year. They’ve handed over the keys, moved the supervisor out of the passenger seat, and skipped the test.

That’s the picture in Licensed to Act: How AI Autonomy Is Outrunning the Limits Placed on It, new research from Illumio and Sapio Research that surveyed 1,100 IT and security decision-makers across 10 markets in August and September 2026.

Seventy percent of organizations now let AI make decisions that needed human approval a year ago. Seventy-six percent have already had at least one AI-related security issue or warning sign. And only 49% can see and audit every action their AI agents take.

Read together, those numbers point to a boundary problem. As an industry, we’ve put most of our energy into telling AI agents how to behave and far too little into defining where they’re allowed to go.

The fix is to plan for failure. Security teams should assume AI agents will eventually do something they shouldn’t, and use segmentation so that mistake can’t spread through the network.

From copilot to driver’s seat: AI agents have started acting

Most of us still picture AI as an assistant that drafts, summarizes, and suggests. The research shows that picture is out of date, especially in the functions security leaders own.

Seventy-seven percent of respondents use AI in IT operations and service management, and 67% use it in cybersecurity operations.

The more telling numbers are about authority. Without any human review, respondents are comfortable letting AI:

  • Run security investigations (32%)
  • Isolate a compromised device from the network (32%)
  • Carry out security remediation (31%)
  • Deploy code changes (29%)
  • Modify production environments (28%)
  • Grant or revoke access permissions (26%)
  • Reset privileged credentials (26%)

Look closely at the bottom of that list. Changing access permissions and resetting privileged credentials are exactly the moves an attacker makes after initial compromise. When a human admin does them, we wrap the work in change tickets and approvals. A quarter of organizations are now comfortable letting software do it alone.

Even the refusals tell a story. The highest “never” rate anywhere in the study was 12%, for letting AI shut down systems. That means nearly 9 in 10 teams would let an agent pull the plug under some circumstances.

In other words, AI already holds the permissions of a senior operator. That raises two questions: How far can an agent go when it acts, and is anyone watching when it does?

Risk appetite is up, and the safeguards are playing catch-up

Security teams like to think of themselves as cautious, but the numbers say otherwise. Sixty-nine percent of decision-makers say their organization is willing to accept higher risk to get the benefits of AI, and 56% accept more AI-related cyber risk than they did a year ago.

Among the organizations accepting more risk, over a third now issue privileged identities to AI. Even worse, 62% say they deployed AI tools before they had the right safeguards in place.

Anyone who was securing infrastructure in the early years of cloud adoption has seen this before. Workloads moved faster than the controls built to protect them. Teams spent years retrofitting security onto environments already running in production, and breaches traveled farther than they should have because nothing inside those environments limited movement.

Plenty of organizations are still cleaning up after their rushed cloud security choices. And now they’re adding AI agents to the mix.

The 35-point gap between AI confidence and visibility

Most security leaders say they’re confident about their AI deployments:

  • 84% say they have clear visibility into how AI systems access data and systems.
  • 81% say they fully understand the risks AI agents introduce.
  • 92% are confident their internal guardrails would stop an AI model from acting outside its role.
  • 92% are confident they could contain a compromised or out-of-role AI system before it reached critical systems or data.

But when the survey asked what security teams can actually observe, the picture shifted. Only 49% can see and audit every action an AI agent takes. Another 32% see most actions, 12% see only outcomes, and 6% have limited or no visibility at all.

That’s a 35-point gap between the visibility organizations claim and the visibility they actually have.

Part of the reason is technical. Most monitoring tools were built for deterministic systems that behave the same way every time they get the same input. An agent chasing a goal can take a different path to the same result on every run.

Seeing only outcomes is a bit like reviewing dashcam footage that shows where the car parked and nothing about how it got there.

Confidence does have a limit, and it’s frontier AI. Nearly half (48%) of decision-makers say they aren’t prepared for the risks the newest, most capable models bring.

Your AI agent is already inside your network perimeter

The consequences are showing up already. Over three-quarters of organizations reported at least one AI-related security issue or warning sign in the past 12 months.

Some of that pressure comes from outside, with 31% reporting attackers who used AI-generated phishing or impersonation to try to get in. But the internal findings are more revealing:

  • 35% saw an AI tool recommend security actions that would have created new risk if followed.
  • 26% had an AI agent connect to systems without security approval.
  • 24% had an agent access data it shouldn’t have.
  • 24% had an agent trigger unauthorized actions.

That first finding deserves a second look. Those risky recommendations were likely caught because a human was still reviewing them.

Almost every other trend in this report points toward taking that human out of the loop. That means today’s incident numbers are likely a floor. As agents move from recommending to executing, the same flawed suggestions become flawed actions.

There’s also a structural reason internal AI is so risky. An attacker has to find a way in. Your own AI agent starts on the inside, already holding valid credentials.

Two-thirds of leaders (66%) say keeping systems patched is no longer enough to prevent compromise, and 61% say blocking known bad behavior has stopped working as a primary strategy.

From a breach containment point of view, an over-privileged agent and a compromised account look the same on the network. Each is an identity with legitimate credentials heading toward places it shouldn’t be. That’s lateral movement, and it’s the problem segmentation solves.

Why guardrails alone can’t keep an agent on the road

When respondents picked their two primary controls for managing what AI can do, three options landed close together. Internal guardrails came in at 44%, external containment controls at 42%, and human oversight at 41%. Explicitly defined permitted actions, or allowlisting, came last at 25%.

That ranking points in a risky direction. Guardrails and human oversight both depend on the AI behaving as it ought to, or on someone catching it when it doesn’t. Yet every incident above involves a system behaving differently from how it was told.

Containment and allowlisting are the only measures on the list that work whether or not the AI cooperates, and allowlisting ranked last.

Think back to the driving example. Road signs and driving lessons shape how a driver behaves. The crash barrier on a mountain pass works whether the driver is paying attention or not. AI agents need both, and right now most organizations have plenty of signs and very few barriers.

The wider industry is landing in the same place. The OWASP Top 10 for LLM Applications 2026 moved excessive agency, which covers AI systems with too much autonomy or too many permissions, up from sixth to third place.

Why segmentation is the answer to AI agent security

The most important finding in this report, in my view, is a contradiction.

Leaders clearly understand that AI changes the game. Eighty-two percent agree AI agents need fundamentally different security controls from traditional software, and 69% agree that limiting what AI can access matters more than trying to predict every threat.

They’re backing that view with money, too. Forty-seven percent are increasing overall security spending, and 42% have set aside new budget for AI-related risk.

Yet when asked which capabilities frontier AI has made more important, only 27% named network segmentation or microsegmentation. That put it dead last of the 10 capabilities assessed. Incident containment barely did better at 30%.

So organizations have correctly identified reach as the thing to control, and then ranked the control that limits reach at the bottom.

My guess is that two old and incorrect assumptions are at work:

  • AI security is mainly a model problem, so the budget flows to the model layer.
  • Segmentation is a long, painful infrastructure project, a reputation left over from the days of virtual local area networks (VLANs) and hand-built firewall rules.

Neither holds up. An AI agent’s risk depends as much on what it can reach as on the model behind it, and reach is exactly what segmentation controls. It defines which systems an agent can connect to and blocks everything else, so a confused or compromised agent can’t move laterally toward critical data.

It also works whether or not the agent follows its instructions, which guardrails and prompts can’t promise. And today’s segmentation doesn’t require re-architecting the network. Policies follow workloads and labels instead of IP addresses and VLANs, so teams can start with their highest-risk agents and expand from there.

The list of everything an agent must never do is endless. The list of what it actually needs to do is short enough to write down, agree on, and enforce. That’s the practical logic of Zero Trust applied to AI, and segmentation is how you enforce it.

How to give AI agents a license with limits

The report closes with six practical steps security teams can take to secure AI agents. Segmentation plays a part in most of them:

  1. Inventory every AI agent, model, tool, and AI-enabled application. Any agent you miss is an agent with no limits at all.
  2. Map what each AI system can access, where it connects, and what it can do. This replaces confidence with evidence, and it gives you the map your segmentation policy is built on.
  3. Give AI agents dedicated identities and least-privilege permissions. Shared or borrowed credentials make every agent look like everyone else.
  4. Define permitted actions on top of prohibited behaviors. Allowlisting is the control that ranked last, but it’s the one that matters most when AI doesn’t cooperate or do what it’s told. Segmentation turns that allowlist into rules the network enforces.
  5. Segment access to critical systems and sensitive data. This puts a boundary between a confused agent and your critical assets, so one agent’s mistake doesn’t become a breach.
  6. Monitor for unusual behavior and contain failures before they spread. Assume something will go wrong. Segmentation is what keeps the blast radius small when it does.

For a deeper look at the identity side of this, my earlier post on why AI agents are becoming digital employees is a good companion read, and our guide to agentic AI security covers the fundamentals.

Why the time to set boundaries with your AI agents is now

It’s tempting to treat AI agent security as a problem for later, especially with the pressure many companies are feeling to get AI implemented as quickly as possible. But the data says that would be a mistake, for three reasons:

  • Autonomy is compounding. Every month, more agents get more identities, more connections, and more permission to act without review. Each one adds to the tangle that someone will eventually have to map and untangle. The cloud era taught us how expensive retrofitting gets.
  • The human safety net is thinning. The most common AI incident in this study was caught by a person in the loop, and organizations are steadily removing that person. When agents act at machine speed, containment has to be in place before the incident starts, because there won’t be time to build it during one.
  • The next wave of AI innovation is already here. Nearly half of leaders say they aren’t ready for frontier AI, and those models will be more capable and more persistent in chasing their goals every day.

The good news is that we don’t need to predict every decision an AI agent might make to control its impact. Instead, we need to decide what it can reach, where it can connect, and how far it can move if something goes wrong.

That’s what segmentation does. It makes AI agent security something you plan for ahead of time, so when an agent goes off course, the damage stays small and the business keeps running.

The UK doesn’t hand new drivers the keys and hope for the best. It sets limits first and loosens them as trust is earned. AI agents deserve the same deal: a license with limits, enforced by the network.

You don’t have to figure out AI agent security alone. Join the Illumiverse, the Illumio user community, where practitioners share what’s working, ask questions, and learn from each other.

Related articles

Experience Illumio Insights today

See how AI-powered observability helps you detect, understand, and contain threats faster.