70% of Enterprises Can't See Their Own Agents

Ask your security team how many AI agents run in your environment right now. Wait for the answer.

If it starts with “we think,” you have a visibility problem. If it’s confident and specific, ask them how they verified it. Ask them to show you the inventory. Most teams can’t. That gap between “we know” and “we can prove” is what researchers call identity dark matter—the stuff you can’t see but can absolutely feel in your risk profile.

Non-human identities already outnumber humans 25-50x in most enterprises, and AI agents add a new, dynamic category that traditional identity tools can’t track. Nearly 70% of enterprises run agents in production, but only 21% have full visibility into their permissions and access patterns. Run an agent census this week: document owner, permissions, last activity, and review date for every agent in production.

The Gap Between Running Agents and Seeing Them

70% of enterprises already deploy AI agents in production. Another 23% are actively building deployments for 2026. But here’s the friction: only 21% of executives report complete visibility into what those agents can access or what permissions they hold.

That’s not a small oversight. That’s a structural blindness.

Non-human identities—service accounts, API tokens, bots—already outnumber human users 25-50x in most enterprises. Add autonomous AI agents to that equation, and you’re dealing with a category that traditional identity and access management tools weren’t built for. Agents don’t behave like static service accounts. They spin up dynamically, modify their own behavior, and operate across multiple platforms and APIs in ways that traditional IAM can’t track.

Native AWS IAM controls you in the AWS ecosystem. Azure AD holds your Azure estate. But when your agent needs to run a workflow that touches Salesforce, GitHub, Slack, and your internal API layer all in sequence, which system is actually governing those interactions? The answer, in most shops, is none of them.

Why Traditional Visibility Tools Fall Short

Your identity and access management platform was designed for humans and static integrations. You provision a user, assign them roles, audit their access quarterly. That works when access patterns are relatively stable.

Agents shatter that assumption. An agent might need different permissions depending on what task it’s running. Some call APIs you didn’t anticipate until runtime. Others authenticate to systems that sit outside your primary IAM platform entirely. Your AWS IAM console won’t tell you what your agent is doing in Jira. Your GitHub audit logs won’t show you what happened when it triggered a deployment workflow.

The tools work within their boundaries. Agents work across them.

A 2026 report from The Hacker News on identity dark matter identified this as a foundational security gap: organizations have visibility into maybe 30-40% of the actual identity traffic flowing through their systems. The rest is blind spots.

The Numbers Tell a Story

CyberArk’s recent research on non-human identity risks found that 87% of enterprises have non-human identities sitting dormant in their systems—accounts created months or years ago that nobody remembers provisioning. Add an agent layer on top of that, and you’re scaling the problem with dormancy, privilege creep, and forgotten owners.

Then there’s the behavior risk. The AIUC-1 Consortium reported that 80% of organizations see risky or unexpected agent behaviors in their production deployments. Actions that shouldn’t have been possible. Permissions that shouldn’t have been granted. But the risky part? Most orgs found out about the behavior only after it happened, from logs, not from real-time visibility.

When you can’t see an agent, you can’t stop it. You can only investigate what it did.

What Visibility Actually Means

More monitoring tools won’t solve this. You need an inventory you can actually trust.

For every agent running in production right now, you need to know:

• Owner. Who created this? Who’s responsible if it goes sideways?
• Creation date and purpose. Why does this agent exist? When did it start?
• Permissions granted. What APIs, systems, and data can it actually access?
• Last activity. When did it last run? Is it still active?
• Review date. When was this agent’s access last verified? Mark one coming up.

That’s a census. Not a scan, not a guess. An actual documented inventory.

Organizations that can’t account for their agents are like cities that don’t know how many people live in them. You can pass all the zoning laws you want, but if you don’t know the population, governance is theater.

Start This Week

Pick a time. Gather your platform teams—the ones running agents in production. Could be ML ops, platform engineering, the AI working group, whoever owns agent deployments at your company. Walk through every agent. Document the five things above. Put it in a spreadsheet or a tool if you have one. Set a quarterly review cycle.

This isn’t a one-time exercise. Agents spawn new agents. Teams ship undocumented deployments. Permissions drift. But once you have a baseline, you have something to defend.

Next step: audit your trust boundaries. Check where agents are authenticating across platforms and whether those handoffs are logged, verified, and governed. That’s the bigger problem—and we’re covering it Thursday.

---