AI agent governance: rules for agents in your company
Agents already act for your staff. Governance decides which ones may, with what access, and who answers when one goes wrong. A policy that no system enforces does not do that job.
Next step
AI Agent Blocklist: the enforcement layer for your agent rules
Turn "unapproved agents are not allowed" into firewall, DNS, browser and MDM rules with a daily list of agent domains and device identifiers.
From $99 a month for one organisation
What is AI agent governance, and how is it different?
AI governance programmes were built for models and their outputs: bias, accuracy, transparency. Agents add a new question. Not "is the answer right?" but "should this software be allowed to do that, with this person's access?"
| Question | Classic AI governance | AI agent governance |
|---|---|---|
| What is governed | Models, training data, outputs | Software that takes actions in your systems |
| Main harm | Wrong, unfair or leaked answers | Wrong actions: sent, bought, deleted, shared |
| Key control | Testing and documentation | Access limits, approval and blocking |
| Who installs it | Data science or IT projects | Any employee, in minutes |
| Evidence | Model cards, test reports | Inventory, policies in force, block logs |
Why enterprise AI agent governance fails in practice
Most programmes do not fail for lack of a policy. They fail because the policy and the real environment drift apart within weeks.
Policy without enforcement
The policy says "approved agents only", but nothing on the network or the devices stops the others.
Approval by email
Requests are approved in a thread. Nobody records the identifier, so nobody can enforce or audit the decision.
Tenant-only view
The team governs agents inside its productivity suite and misses extensions, AI browsers and desktop agents.
No owner per agent
Agents outlive projects and staff. Without a named owner, nobody reviews or retires them.
One-time review
A review at rollout, then never again, while the agent gains features and permissions with every update.
Lists that go stale
A hand-made block list of ten agents is out of date within a month. New agents appear every day.
An AI agent governance framework in seven layers
Each layer produces something you can show an auditor. If a layer has no artefact, it is not working yet.
Policy
States which agent types are allowed, which are banned and how to ask for a new one. Short enough that staff read it.
Inventory
Lists every agent found, by identifier, with owner and reach. Re-matched weekly against a current agent list.
Approval
Every agent passes the same review: what it does, what it reaches, where data goes, what it costs.
Identity and access
Approved agents get their own accounts or scoped tokens, the fewest permissions that work and allowed sites only.
Enforcement
Everything not approved is blocked by domain on the network and by identifier on devices and in browsers.
Monitoring
Watches for new agents, blocked attempts, new consents and unusual activity from agent accounts.
Review and retire
Every approval has an end date. Owners renew it or the agent is blocked and its tokens are revoked.
Layers 2 and 4 have their own guides: AI agent inventory and agent identity and permissions. For layer 3, use our AI agent risk assessment.
Frameworks your agent governance can map to
You do not need a new framework. Map the seven layers to the ones your auditors already know. Check the current text of each, as several are being revised.
January 2023
NIST AI Risk Management Framework 1.0
Four functions: Govern, Map, Measure and Manage. Agent inventory and approval sit naturally under Map and Govern.
July 2024
NIST AI 600-1, Generative AI Profile
A companion profile for risks specific to generative AI, with suggested actions you can apply to agents.
2023
ISO/IEC 42001
The management system standard for AI. Your agent policy, inventory and reviews become controls inside it.
In force since August 2024
EU AI Act
Regulation (EU) 2024/1689. Its obligations phase in over several years, so check which apply to your use of agents and when.
January 2026
Model AI Governance Framework for Agentic AI
Published by Singapore's Infocomm Media Development Authority as guidance for organisations deploying agents.
December 2025
OWASP Top 10 for agentic applications
A security risk list for agents. Useful for the threat side of your risk register; see AI agent security risks.
Analyst advice: in December 2025 Gartner published "Cybersecurity Must Block AI Browsers for Now". Many teams cite it to justify a default-deny rule for AI browsers. More on our page about Gartner's advice to block AI browsers.
Who does what: a simple ownership table
R means responsible for doing it, A means accountable for the result, C means consulted. One accountable person per row keeps decisions from stalling.
| Activity | CISO or CIO | Security team | IT admins | Legal and privacy | Business owner |
|---|---|---|---|---|---|
| Write the agent policy | A | R | C | C | C |
| Keep the inventory | A | R | C | ||
| Approve a new agent | A | C | C | R | |
| Set accounts and access | A | R | C | ||
| Enforce blocks | A | R | |||
| Review data use and contracts | C | A | R | ||
| Recertify each agent | A | C | R |
Governance switches in the platforms you already run
Vendors now ship agent controls in their admin consoles. Use them, and know where each one stops. Menu names change often, so check the current admin console.
| Platform | Governance switch | What it does not cover |
|---|---|---|
| Microsoft 365 admin center | Agents > Settings: Agent and plugin access, User access, Sharing and Agent management rules such as blocking ownerless agents without usage | Agents installed on devices or in browsers outside the tenant |
| Google Admin console | Devices > Chrome > Apps & extensions: block extensions by ID, or block all and allow a list | Other browsers, desktop agents, network traffic |
| Windows | Experimental agentic features stays off by default and only an administrator can turn it on; newer WindowsAI policies control agent connectors | Third-party agents and AI browsers |
| Business AI workspaces | Workspace admin settings for agent features, site lists and roles in the vendor's own console | Personal accounts of the same product on the same laptop |
| Network | DNS filter, firewall or secure web gateway with custom or external lists | Nothing on its own about which device app made the request |
Step-by-step guides: Copilot actions and Microsoft 365 agents and Chrome AI policies in the Google Admin console.
Check the defaults, not just the switches
Agent features often arrive switched on. Governance has to start with an audit of what your vendors enabled for you.
Tenant agents
Microsoft states that agent management in the admin center is enabled by default in Copilot-licensed tenants, and User access defaults to All users.
Browser agent extensions
Some business AI plans turn their browser agent on for the whole team by default until an owner switches it off.
Operating system agents
Windows keeps its experimental agent setting off by default. Re-check after every feature update.
The pattern: every vendor governs its own agents well. Nobody's console governs the agents of every other vendor. That cross-vendor gap is where a shared block list does its work.
An approval route staff will actually use
If asking takes weeks, people install agents anyway. A short, predictable route keeps requests coming to you instead of around you.
Request
One form: which agent, for what task, which accounts and data it needs, who owns it.
Triage
Look the agent up in your reference list. Its category and risk grade decide whether it gets a light or a full review.
Review
Security, privacy and the owner answer the same questions every time: reach, data flows, actions, cost, exit plan.
Set up
Create the agent account or scoped token, allowed sites and confirmation rules before anyone uses it.
Publish
Add the agent to the approved list, lift the block for the right group and tell staff it is available.
What to report to leadership each month
Six numbers tell a board whether agent governance works. Each comes from tools you already run once the inventory and blocks are in place.
Coverage
Agents in the inventory
Total found, split into approved, monitored and blocked. A rising blocked share early on is normal.
Accountability
Approved agents with an owner
Should be all of them. Any gap is an action item with a date.
Enforcement
Blocked attempts per week
From DNS, firewall and browser logs. Shows where demand is, and which teams need an approved option.
Freshness
New agents found this month
Shows the inventory is being re-matched, not left to age.
Hygiene
Reviews overdue
Approved agents past their review date. Overdue means blocked unless renewed.
Service
Days to decide a request
If this grows, shadow agents grow with it. Keep the approval route fast.
Where are you today? A five-level maturity ladder
Unaware
No policy, no list. Agents run wherever staff install them.
Written
A policy exists. Nothing enforces it.
Visible
An inventory exists, matched against a current agent list.
Enforced
Unapproved agents are blocked on network, browser and devices.
Managed
Approved agents have owners, scoped access and review dates.
Most organisations we talk to sit at level 1. Moving to level 3 is mostly a matter of feeding a current agent list into tools you already own.
Do not skip a level. Enforcement without an inventory blocks tools people rely on and creates angry tickets. An inventory without enforcement documents risk without reducing it. Level 4 only works once the first three are routine.
Eight policy clauses to adopt now
- Staff may use only AI agents on the approved list, on work devices and work accounts.
- AI browsers and agent browser extensions are blocked unless approved for a named role.
- Agents may not be connected to mail, files or business systems without a security review.
- Every approved agent has a named owner and a review date no later than twelve months out.
- Agents use the fewest permissions that work and, where possible, their own accounts.
- Sending, buying, deleting and publishing by an agent require a person to confirm.
- MCP servers and agent add-ons are installed only from the approved list.
- IT may block, remove or revoke any agent that breaks these rules, without notice.
A full template with explanations is in our AI agent acceptable use policy guide. Deciding between a default-deny and a default-allow stance is covered in allowlist or blocklist for AI agents.
How we solve it: the enforcement layer
The AI Agent Blocklist turns clauses 1, 2 and 7 into rules your existing tools apply. Every entry is checked before it enters the list, and the lists update every day, so the rules keep pace with new agents.
Network
2,600+ agent domains for DNS filters, firewalls and gateways, as plain lists, hosts, PAC and external list files.
Browsers
Chrome and Edge extension block lists for Group Policy, Intune and the Google Admin console.
Devices
macOS bundle IDs and Windows package IDs for Jamf and Intune, plus process names.
Developers
Command-line coding agents with their package names, for package and endpoint rules.
| Governance need | How the list helps |
|---|---|
| Default-deny for agent categories | Block whole categories: AI browsers, autonomous agents, agent platforms, MCP servers, coding agents |
| Risk-based rules | Use the risk grade on each entry to block higher grades first and monitor the rest |
| Approved exceptions | Remove approved tools from your copy of the list, or allow them above the block rule |
| Audit evidence | Policies in force plus block logs from your own tools show enforcement |
| Keeping current | Daily updates bring new agents into rules without manual research |
For MCP servers in particular, pair the block list with our MCP security checklist. To see what is hiding before you enforce, start with shadow AI agent discovery.
Make your agent policy enforceable
- 2,600+ AI agent domains and 5,300+ device identifiers
- Tool name, category, identifier type and risk grade on every entry
- Formats for firewalls, DNS filters, gateways, browser policy and MDM
- Updated every day; download access set up within one business day of your order
Full file list: the AI Agent Blocklist.
Ask us or order a plan
Tell us what you need to block and where your rules live. We reply within one business day.
- Files for firewalls, DNS filters, secure web gateways, MDM and browser policies
- Updated every day as new AI browsers, agents and extensions appear
- Plans for one organisation, for MSPs and for product makers, or a hosted filter we run for you
AI agent governance: FAQ
What is AI agent governance?
It is the set of rules, roles and controls that decide which AI agents may run, for whom, with what access, and who answers for them. It covers policy, inventory, approval, identity and access, enforcement, monitoring and review.
What is an AI agent governance framework?
It is the structure you use to run agent governance: named layers, owners and evidence for each. Many organisations map it to the NIST AI Risk Management Framework, ISO/IEC 42001 and, in the EU, the AI Act. Singapore published a model framework for agentic AI in January 2026.
What is an AI agent governance layer?
It usually means the technical layer that applies governance rules at run time: what may be installed, which domains agents may reach, which accounts they use and which actions need approval. Without it, a policy is only a document.
What does Gartner say about AI agents and browsers?
In December 2025 Gartner published advice titled "Cybersecurity Must Block AI Browsers for Now". Many security teams use it as the starting point for a default-deny rule on AI browsers and agent extensions.
How does Microsoft handle AI agent governance?
For agents in a Microsoft 365 tenant, the admin center has Agents pages with a registry, settings for which agent types users may install, who may access and share agents, and rules such as blocking ownerless agents. Agents installed outside the tenant need device and network controls.
Which tools do we need for AI agent governance?
Mostly ones you own: MDM or endpoint tools, browser policy, a DNS filter or firewall, an identity provider and logging. What is usually missing is a current list of agents to feed into them.
Who should own AI agent governance?
One accountable executive, usually the CISO or CIO, with security writing the rules, IT enforcing them, legal and privacy reviewing data use, and a named business owner for every approved agent.
Microsoft, Microsoft 365, Windows, Intune and Edge are trademarks of Microsoft Corporation. Google, Chrome and Google Admin are trademarks of Google LLC. Jamf is a trademark of Jamf Software, LLC.
Gartner is a trademark of Gartner, Inc. OWASP is a trademark of the OWASP Foundation. ISO and IEC are trademarks of their respective organisations.
Names are used only to identify the products and organisations discussed. AI Agent Blocklist is not affiliated with, sponsored or endorsed by these companies.