Start with risk categories, not tools
Begin the gap check from the risks an AI agent introduces, then ask which tools contribute to each one. Starting from the tool list tends to confirm what each product already does; starting from risk categories shows what the stack cannot see. The check uses five categories.
| Risk category | What it asks |
|---|---|
| Agent identity and permissions | Who or what is acting, and with which delegated rights. |
| Data exposure through prompts and context | What sensitive data reaches models and tools. |
| Tool and MCP access | Which tools and connections an agent can reach and use. |
| Runtime policy enforcement | Whether rules apply at the moment of action. |
| Auditability | Whether actions can be reconstructed and evidenced. |
Define what full, partial and absent mean
Rate each combination of tool and risk category as full, partial or absent, and base the rating on evidence from your own environment rather than on vendor descriptions. The deciding question is what the tool can observe and what it can enforce for an agent. A tool that can only observe an action offers different coverage from one that can enforce policy on it.
Use a simple matrix to record the results. Rows are the risk categories and columns are the existing control classes.
| Risk category | CASB | DLP | IAM | EDR | SIEM | API gateway | AppSec |
|---|---|---|---|---|---|---|---|
| Agent identity and permissions | Rate | Rate | Rate | Rate | Rate | Rate | Rate |
| Data exposure through prompts and context | Rate | Rate | Rate | Rate | Rate | Rate | Rate |
| Tool and MCP access | Rate | Rate | Rate | Rate | Rate | Rate | Rate |
| Runtime policy enforcement | Rate | Rate | Rate | Rate | Rate | Rate | Rate |
| Auditability | Rate | Rate | Rate | Rate | Rate | Rate | Rate |
What to test for each existing control class
The classes in the matrix (CASB, DLP, IAM, EDR, SIEM, API gateways and AppSec tools) were built around human users, static applications and known data flows. Agents act differently, so test each class against each risk category with the same questions:
- What can the tool observe about an agent's actions?
- What can the tool enforce, as opposed to only monitor?
- Can the evidence be produced from your own environment, rather than assumed from product documentation?
Decide between extending a tool and adding runtime controls
Once the matrix is filled in, sort each gap into one of two groups.
Gaps you can close by extending a tool
The first group can be closed by extending a tool you already operate: adding agent events to the SIEM, issuing separate service identities in IAM, or tuning DLP policies for new data paths. Extension is usually cheaper and keeps operations in familiar workflows, so it should be the default where the tool can enforce rather than only observe.
Gaps that depend on the runtime layer
The second group contains gaps that depend on the runtime layer itself. These include attributing each action to a distinct agent identity, applying least-privilege permissions per agent, approving or denying individual tool calls and MCP connections as they happen, and producing an audit record of what the agent did and under whose authority. If a tool's architecture cannot see or intercept these actions, tuning will not close the gap.
Criteria for evaluating any new AI runtime control
If the gap check points to a runtime layer, evaluate any candidate against these criteria.
- Gives each agent a distinct identity and attributes every action to it, including the human or system it acts for.
- Enforces least-privilege permissions per agent and per tool, with the ability to revoke them.
- Applies policy to tool calls and MCP connections at runtime, with enforcement and not monitoring alone.
- Supports approval workflows for sensitive tools or actions.
- Sends structured logs to your existing SIEM and works with your IAM, without parallel identity stores.
- Produces audit records that show who, what, when and under which policy, suitable as compliance evidence.
Make the gap check repeatable
Keep the matrix, the rating for each cell and the evidence behind it together, so the same check can be run again as agents, tools and permissions change. Using the same five risk categories and the same full, partial or absent ratings each time makes results comparable and shows whether extending tools or adding runtime controls has actually closed a gap.