Skip to main content

Business pressure to adopt agents is outpacing the security team’s ability to evaluate them.

Either security says no, and becomes the thing standing in the way of the strategy, or the business moves ahead anyway, and the agent estate expands unmanaged.

But there’s a third way.

2026: The year the enterprise gained a new labor ecosystem

Four key challenges in controlling agents

  • 01

    You can only secure what you know

    Most teams cannot compile a complete, current list of what agents are deployed, where they operate and who owns them, so posture, compliance and control decisions are all made on incomplete data.

    327% more agents than one customer’s own security team estimated

  • 02

    Audit logs are optional, and lack behavioral context

    Not every agent produces logs, and those that do often allow the session to be rewritten. Without behavioral context, a risky session reads like routine work.

  • 03

    Agents span surfaces. Controls sit on one.

    Endpoint-only sensors stop at the device, and gateways see only the traffic that crosses them. Neither can govern a system that lives everywhere at once.

  • 04

    Risk is financial too, and it compounds

    Developer agents wasting tokens on side projects. Agents that keep running long after they should have stopped. Three departments paying separately for the same work. It comes up in every other first meeting, and often before security does.

1. You can only secure what you know

Every organization we have deployed into found more agents than it expected, in one case 327% more. The gap between the inventory and the estate is where risk, spend and accountability all go missing at once.

In the register reviewed, owned, approved

Running anyway shadow, citizen-built, developer preference, duplicated across departments

Unapproved customer-facing agents processing sensitive complaints. Developer agents running side projects on company tokens. A marketing agent built by someone who never knew a review process existed. None of it appears in the register, and none of it stops running while the register is being written.

  1. 09:14:02 Repository read · looks like development
  2. 09:14:38 Credential used · looks like development
  3. 09:15:11 Commit pushed · the credential belongs to a competitor

Without behavioral context, all three rows are the same row.

2. Routine and risky look identical in a log line

Many agents leave no log at all, either because the platform they run on doesn’t produce one or because the security team has no access to it, and where a log does exist, the session can often be edited or rewritten after the fact.

Even a complete log only records what happened, one event at a time. It shows that a coding agent read a repository, used a credential and pushed a commit, but not the full picture. Understanding intent means reading the agent itself: its objective, its context and the sequence of what it did.

3. Agents cover multiple surfaces. Most tools just watch one

Which surfaces each category of tool covers, compared with the surfaces an agent actually spans
Tool CloudCodeEndpointBrowser
Total agent coverage covered covered covered covered
Endpoint sensor not covered not covered covered not covered
MCP gateway covered not covered not covered not covered
Cloud posture tool covered not covered not covered not covered
Identity platform permissions only, not activity permissions only, not activity permissions only, not activity permissions only, not activity
Cost dashboard not covered not covered not covered not covered
  • covered
  • permissions only, not activity
  • not covered

The problem is that the agent is a single system and each tool is looking at one wall of it. Assembling five partial answers by hand, weekly, gives you a history of the estate rather than a working view of it.

4. Risk is financial too, and it compounds

Agentic spend is now a board-level line item. Wasteful spend can come in many forms: side projects, looping agents, duplicate work, and liberal model choice. Security, AI operations and finance teams need a shared view of not just what agents cost, but what work that spend is actually buying.

What you get today

Platform invoice · last month

API and model usage $418,900

One aggregate, with no agent, owner or task attached, and no answer to “was any of that worth it?”

What the question actually needs

The same month’s spend, attributed to the agent that spent it
Agent Work it was doing Owner Spend
claims-triage Claims intake Claims ops $96,400
market-research Competitor research Strategy $61,200
pr-review-agent Retrying a failed job, 3,400 times Platform eng $73,800
unnamed-dev-agent Personal side project Unattributed $44,100
3 × content-drafter The same task, three departments Marketing, Sales, CS $143,400

Three of these five rows are waste, and none of them are visible until spend is attached to behavior.

The security, cost, and activity record are the same record

The security record

Which agents exist, what they can reach, what they did, how risky it was.

Owned by security

The financial record

How many tokens, on which model, in which month, at what total.

Owned by finance or IT

The activity record

What business outcome the agent achieved.

Owned by the business department

Often treated as three separate records, but the same agent activity underlies all three of them.

Six real examples of risks we find in customer environments

  • Critical Supply chain

    An agent connected to a malicious skill

    It looked legitimate until somebody understood what the agent was doing with it. Skills sit in the agent’s own configuration and generate no outbound traffic for a gateway to intercept.

  • High Data exposure

    An unapproved agent processing sensitive complaints

    Customer-facing, never reviewed, handling exactly the material a review would have caught. It was not in any inventory.

  • Critical IP leakage

    Developer agents pushing corporate secrets to public repos

    In the commit history it reads as routine work. The distinction only exists if you can see what the agent was holding at the time.

  • Medium Propagation

    A worm propagating via agent hooks

    Agent-to-agent spread through the same mechanisms that make agents useful. Nothing in the traditional stack is watching that path.

  • High Insider threat

    A coding agent running on credentials linked to a competitor

    An employee ran an unsanctioned agent with credentials linked to a rival, exposing customer data and company IP. Endpoint and network controls saw normal development.

  • Medium Financial risk

    Developer agents burning tokens on side projects

    Spend runs without an owner, attribution or a cap. On the invoice it looks like a usage spike, and it only reads as a security problem once you can see whose credentials paid for it.

The window to act is now, while the infrastructure is still forming

  • 33%

    of enterprise software applications will include agentic AI by 2028

    Gartner

  • 15%

    of day-to-day operations completed autonomously by 2028

    Gartner

  • 40%

    of agentic AI projects will face cancellation due to unmanaged risks

    Gartner

  • 12

    months in which every major platform shipped agent capability

    Anthropic, Microsoft, OpenAI, GitHub, AWS

Closing the agent governance gap starts with getting closer to the agent

Solving agentic security starts with an understanding from inside the agent’s own execution path, not from a boundary near it. From there, a security team can approve more agents, with evidence behind every decision.