AI Governance Glossary

Working vocabulary for AI agent governance, MCP security, and policy enforcement.

A

3 terms

Adaptive Policy Layer #

Policy Maturity: top tier

An adaptive policy layer combines expressive rules, session-based controls, and intent-based outcomes as one system, closing gaps as new tools or integration paths emerge rather than requiring a rewrite each time the environment changes. It’s the compounding version of policy maturity: each layer catches what the others can’t, and the system gets more effective the more it’s used.

Agent Identity #

Agent identity treats an AI agent as a distinct actor with its own traceable record of actions, rather than an extension of the human who deployed it. This matters for audit and incident response specifically: “which agent did this, under which policy, with what justification” is a different — and harder — question than “which employee’s credentials were used.”

Allow / Deny / Redact #

Allow, deny, and redact are the three enforcement outcomes a policy layer can apply to any agent action: allow the action to proceed, block it entirely, or let it proceed with sensitive content removed or masked. Redact matters specifically because many legitimate agent actions involve a mix of permitted and sensitive data — blocking the whole action is often the wrong answer when removing one field would have been enough.

B

1 term

Blast Radius #

In the context of AI agents, blast radius is the scope of damage possible if a given agent’s access were misused or compromised — not what the agent is intended to do, but the worst case of what it’s technically able to do. Asking “what’s the blast radius if this goes wrong” before granting access reframes agent permissioning as a worst-case question rather than a best-case one.

C

1 term

Control Plane #

Key concept

A control plane is the layer that defines and enforces the rules governing a system, as distinct from the “data plane” that carries the actual traffic. For AI agents, a security control plane decides what each agent can access, under what conditions, before any tool call executes — it’s the governance logic, not the pipe the traffic flows through.

E

3 terms

Excessive Agency #

Key concept

Excessive agency is the risk that an AI agent has more access, permissions, or ability to take action than its task actually requires — named by 600+ security experts as a top AI deployment risk. It’s the agent-era version of the principle of least privilege: an agent built to summarize emails that also has the ability to send them, delete files, or call payment tools carries excessive agency even if it never misuses that access.

Expressive Policy #

Policy Maturity: Level 1

An expressive policy is a rule written in plain, explicit language that mirrors how a security team already thinks and talks about risk — e.g., “block payment tool calls from users outside Finance.” It’s the entry point to a policy layer: readable enough to write without AI security expertise, specific enough to enforce immediately.

Expressiveness & Usability #

These are the two criteria for evaluating any AI governance tool: expressiveness is whether the policy model can actually represent the rule a security team wants, with no exceptions or workarounds; usability is whether someone without AI security expertise can write and maintain that rule. Most vendor feature lists — logging, alerting, dashboards — ultimately serve one of these two criteria in disguise.

I

1 term

Intent-Based Policy #

Policy Maturity: Level 3

Intent-based policy enforces the outcome an organization wants — “don’t let customer financial data leave the organization” — rather than enumerating every rule needed to produce that outcome. Because it’s evaluated against what’s actually happening rather than a fixed rule list, it survives changes to tools, vendors, or integration paths that would otherwise require rewriting every rule by hand.

M

1 term

MCP Gateway #

Key concept

An MCP gateway is the control point that sits between AI agents and the tools, data, and systems they’re allowed to call through the Model Context Protocol. Every tool call passes through it, so it’s where authentication, authorization, and policy enforcement happen before a request reaches a real system — rather than relying on each agent or tool to enforce its own rules. Dtwo operates as this layer, applying policy on every action in real time.

N

1 term

Non-Human Identity #

Key concept

Non-human identity refers to treating AI agents, service accounts, and automated systems as first-class identities in their own right — with their own credentials, scoped permissions, and audit trail — rather than inheriting the access of the human who deployed them. Traditional identity and access management was built around human users; agents that act autonomously need an identity model that can be scoped, monitored, and revoked independently.

P

3 terms

Policy Catalog #

A Policy Catalog is an open, shared library of pre-written, enforceable policies that organizations can adopt, adapt, or contribute to rather than authoring every rule independently. Treating policy as a shared, improvable resource — rather than something each security team solves in isolation — is what makes a Policy Catalog different from vendor documentation: it’s meant to be contributed to, not just consumed.

Policy Kit #

A policy kit is a pre-built set of policies designed for a specific industry, team, or function (e.g., RevOps, Healthcare, Engineering) that extends existing human permissions to agents rather than requiring policy to be written from scratch. Policy kits exist to solve the blank-page problem — most organizations already know what a human in a given role shouldn’t be able to do; a kit translates that into enforceable agent controls without each team reinventing it.

Policy Layer #

The policy layer is the governance logic that decides what each AI agent is allowed to do — distinct from the security gateway that routes traffic and the agents that execute tasks. Most organizations already have a security gateway and already have agents; what’s missing is a policy layer sophisticated enough to govern what dynamic agents actually do once deployed, rather than just what they’re connected to.

R

1 term

Reachability vs. Meaning #

Reachability governs which tools and systems an agent is technically permitted to connect to; meaning governs what a given action actually does once the agent is inside — evaluating each call by its arguments and by what the session has already done. A system can get reachability right (the agent is only connected to approved tools) and still miss real risk if it doesn’t also judge the meaning of each call.

S

2 terms

Session-Based Policy #

Policy Maturity: Level 2

A session-based policy responds to context and sequence within a single agent session rather than evaluating each tool call in isolation — e.g., blocking outbound email only if the agent has already accessed financial records earlier in that same session. This catches risks that no single-action rule would flag on its own, because the risk lives in the sequence, not any one step.

Shadow AI #

Key concept

Shadow AI refers to AI tool usage happening outside an organization’s sanctioned, governed, and audited systems — employees pasting data into ungoverned models, agents connecting to tools without IT’s knowledge, or AI usage with no contract, audit relationship, or ability to revoke access. It’s a governance blind spot rather than a malicious one: most shadow AI use isn’t adversarial, it’s just invisible to security teams until something goes wrong.

Get in touch

Tell us what you need to govern.

A few lines is plenty — what you're deploying, and what you need control over. We read every one.