How We Secure Our AI Agents: The Rules Every Agent at ITMTB Follows

An agent is not an app and not a person. The rules we apply to every agent that works on our platform, what we refuse to let agents do, and a checklist you can use on any vendor.

How We Secure Our AI Agents: The Rules Every Agent at ITMTB Follows

Most advice on how to secure AI agents is written for companies still thinking about agents. This post covers the agents that already work on our platform, Tomorrow Central.

The short answer: every agent on our platform acts for a named person and borrows that person's permissions, never more. A person grants access on a consent screen and can withdraw it at any time. Agents read freely within that grant, but a change runs only if the organisation's rules name it and it has been approved, by those rules or by a person. Our access to a customer's cloud can only read. What an agent reads is treated as data, never as instructions. Every call is checked on our side and recorded.

For CTOs and heads of engineering deciding whether to let agents near their systems: the nine rules we follow, what we refuse to let agents do, and a checklist for any vendor, including us.

Why do AI agents need different security from apps or people?

Security models were built for software, which does what its code says, and for people, who you vet and give a role. An agent is neither.

  • It decides its own next step. You cannot test every path an agent will take, because it chooses the path at run time.
  • It acts at machine speed. A person who is about to do something odd usually pauses. An agent does the odd thing a hundred times before anyone looks.
  • It can be steered by what it reads. A web page, a document or a resource name can contain text written to look like an order. People ignore that. Models sometimes don't.
  • It acts for someone. The real question is never "what may this agent do?" but "what may this person do, and how much of that did they hand over?"

OWASP calls the resulting risk excessive agency, and its remedies include minimal permissions, a person's approval for high-impact actions, and acting in the user's own context. Our rules turn those ideas into mechanisms. For the wider governance picture, see our analysis of AI agents and enterprise cybersecurity.

The nine rules every agent follows

1. An agent borrows a person's permissions, never more

What it means. An agent has no master key of its own. It acts for one named person, and what it can do is the overlap of four things: what that person may do, what their organisation has switched on, what the person approved for this agent, and what the agent asked for. If any one of those says no, the answer is no.

Why. It ends the abstract argument about what agents "should" do: the question starts with whether the person may. If the organisation has not switched a product on, no agent of theirs can reach it. When a permission lookup fails, the answer is no, never "unrestricted".

The pass the agent carries names both parties: the person it acts for and the agent doing the acting. This is delegation in the sense of the OAuth token exchange standard, RFC 8693: the agent acts for the person and keeps its own identity, so every record shows who asked and who acted.

2. Access starts on a consent screen, and changes start unticked

What it means. A person connects an agent by signing in and approving a consent screen that names the agent and lists each permission it wants as a tick box. Reading starts ticked for a recognised agent; anything that changes something starts unticked. An agent that registered itself gets nothing ticked in advance. Unticking a box withholds that permission.

Why. A yes that a person never actually gave is not consent. A pre-ticked "allow changes" box is a trap with good manners.

3. Consent can be withdrawn, and withdrawal ends access

What it means. Every connected agent appears in the person's dashboard with a disconnect button. Disconnecting does two things: the agent's current access stops being accepted, and it cannot renew that access. From then on it can start nothing new and read nothing back. Work it had already started, with permission, runs to the end.

Why. If withdrawing takes a support ticket, people stop granting. And withdrawal must end the grant itself, not just the current pass, or the agent simply asks for a new one.

4. Access passes are short-lived, work only with us, and are checked on every call

What it means. The pass an agent receives is valid for a short time, is renewed only while the person's grant still stands, and is issued for our platform only, so it is useless anywhere else. When a person is deactivated or removed from their organisation, their agents can no longer renew access either. Every request is checked on our side, before any tool runs; nothing relies on the agent behaving well.

Why. A short life and a single audience limit what a leaked pass is worth. The MCP authorisation specification asks for the same things: OAuth, tokens that a server checks were issued for it, and short-lived access tokens.

5. Reading is free within the grant; changing needs approval

What it means. Every action we offer agents is labelled as reading or as changing something. Reading runs within the person's grant. A change made by an agent goes through an approvals step first. The organisation's rules decide which changes may happen at all, and a change that no rule names is refused. The same rules decide which permitted changes run at once and which wait for a person to say yes. Either way, the approval covers one action with its exact details, once, and the system that carries out the change checks for it before acting.

Why. An agent that picked up the wrong file acts on it as quickly as on the right one. Checking the approval where the change runs, not only where it was asked for, leaves no side door.

6. Our access to your cloud can only read

What it means. When a customer connects a cloud account, they create a role in their own account that our platform can use. That role can list, describe and read configuration and cost data. It cannot create, change, stop or delete any of the customer's resources. It only works when presented with a code that is unique to that connection, so nobody else can borrow it. Only a person can create that role, so an agent cannot link a new cloud account on its own.

Why. The safest way to stop an agent changing your cloud is not to give it the ability. Our findings are advice with evidence; a person decides and acts.

7. What an agent reads is data, never instructions

What it means. Resource names, tags and page text are written by someone, possibly an attacker. Anything that came from a scanned account reaches the agent inside a labelled envelope that says it is untrusted content to report on, not commands to follow. Our tool instructions say the same in plain words.

Why. OWASP's guidance on prompt injection is to separate and clearly mark untrusted content, keep privileges minimal and put a person in the loop for privileged actions. Labelling alone will not stop every injection, so an instruction that gets through still meets borrowed permissions, read-only cloud access and an approvals step.

8. Every tool says plainly whether it changes anything

What it means. An agent decides whether to call a tool from its description. So every tool we publish declares a title and states, explicitly, whether it only reads; our automated tests fail if a new tool skips that decision. Server instructions are kept short enough for a client to read them in full. Errors carry a stable code and the action that clears it, so an agent does not guess its way into a retry loop.

Why. A wrong "read-only" label never throws an error; it just misleads. Testing descriptions is the only way to keep them honest.

9. Secrets are stored unreadably, every call is traced, and we use it first

What it means. Keys, client secrets and renewal tokens are stored only as one-way hashes, and keys and secrets are shown once. The key that signs every pass is held in a cloud key-management service and never leaves it, not even to us. Every call carries a correlation id that follows it across our services, secrets are redacted from logs, and each agent call is recorded with the person, the organisation and the agent behind it. And our own company is a tenant on the platform: our tools run on our own website and data before they run on a customer's.

Why. A stolen database of hashes holds nothing usable. Tracing every call means "which agent did this, and for whom?" always has an answer. And the first to feel a bad default is us.

That is what we mean by secure by design: agents see only what you allow, and change only what your rules and your people approve.

What we deliberately don't let agents do

Some limits are not settings; they are absent by design:

  • Delete. None of the actions we offer agents deletes a project, a website or a cloud resource. An agent can be wrong confidently and fast, and a deletion cannot be taken back. The few actions that overwrite something, such as re-importing a catalogue or replacing a setting, are changes, so they need approval first.
  • Change your cloud. Our cloud access is read-only. An agent can tell you what is idle and what it costs; it cannot stop or remove it.
  • Hand out new trust. An agent cannot link a new cloud account on its own, and it cannot widen its own permissions. More access means a person approving it again.
  • Exceed the person. No setting lets an agent do more than the person it acts for, or reach a product their organisation has not switched on.

A checklist for any AI agent or vendor

Use this on any agent, including ours. Ask the vendor to show each answer, not describe it.

AskA good answerA warning sign
Whose permissions does the agent use?The approving person's, narrowed to what they granted and what the organisation allowsA shared service account or an admin key "for convenience"
What does the consent screen show?The agent's name and each permission in plain words; changes unticked by defaultOne "connect" button, or permissions you cannot untick
How do I withdraw access, and when does it stop?A button in your own dashboard; current access refused and no renewalA support ticket, or "when the token expires"
How long does a pass last, and where is it valid?Short-lived, issued for one service, renewed only while the grant standsLong-lived keys accepted everywhere
Where are permissions checked?On the vendor's side, on every request, before the tool runsIn the agent's prompt or the client
What needs approval?Every change, by your rules or a person; the approval covers one action, once, and is checked where the change runs"The agent asks if it's unsure"
What access to my cloud or data does it need?Read-only, guarded by a code unique to your connectionWrite or admin roles for a reporting job
How is untrusted content handled?Labelled as data; narrow permissions and approvals behind it"The model is trained to ignore injections"
Where do secrets live?Hashes only; the signing key in a key-management serviceEncrypted, recoverable secrets in a database
Can I trace one call end to end?A correlation id naming the person, organisation and agentLogs per service with no join
Do tools say whether they change things?Every tool labelled, and the labels testedLabels missing or never checked
Does the vendor use it on itself?Yes, on its own systems, before customersOnly in demos

A vendor that fails three or more of these is asking you to trust the model. Our rules exist so you don't have to.

Key takeaways

  • An agent should borrow a named person's permissions, narrowed by what that person approved and what the organisation allows.
  • Consent belongs on a screen with real tick boxes, and withdrawing it must stop both current access and renewal.
  • Reading can run freely within the grant; every change needs a rule that names it and an approval checked where the change runs.
  • Read-only access to outside systems is the strongest defence against prompt injection, because an injected instruction has nothing to act with.
  • Ask vendors to show their answers, not describe them.

Questions people ask

How do you secure an AI agent?

Give it no authority of its own. Let it act for a named person, within what that person approved and what their organisation allows; check every call on the server, not in the agent; keep access to outside systems read-only where you can; and clear every change through approval rules before it runs.

What permissions should an AI agent have?

No more than the person it acts for, cut down to what that person approved for that agent, and further limited by what their organisation has switched on. The agent should never hold a master key of its own.

How do you stop prompt injection in AI agents?

You cannot rely on the model to ignore it, so limit the damage instead: label everything read from outside as data, keep the agent's permissions narrow, make access to outside systems read-only where possible, and require approval before any change.

Is MCP secure enough for business use?

The protocol's authorisation rules are sound: OAuth, and tokens issued for one server that the server must check are meant for it. Whether a particular MCP server is safe depends on how it implements consent, permission checks, logging and approval for changes, so ask the vendor for each.

Can an AI agent delete data or resources?

Not on our platform. None of the actions we offer agents deletes a project, a website or a cloud resource, and our access to a customer's cloud is through a role that can only read. The few actions that overwrite something, such as re-importing a catalogue, are changes, so they go through the organisation's approval rules first.

How do I withdraw an agent's access?

On our platform, disconnect the agent from your dashboard. Its current access stops being accepted and it cannot renew it, so it cannot start anything new or read anything back.

What should I ask an AI agent vendor about security?

Whose permissions the agent uses, what the consent screen shows, how access is withdrawn and how quickly, whether changes need approval, what access to your cloud looks like, how untrusted content is handled, where secrets live and whether every call is logged and traceable.

Where to start

To see these rules from the agent's side, read how agents connect to Tomorrow Central, or connect your own agent from the Tomorrow Central home page. For what the platform does, see Tomorrow Central. If you are building or deploying agents of your own and want them threat-modelled, scoped and tested before they go live, see our cybersecurity services, which include AI agent security architecture reviews, or contact us.

Sources

  • OWASP Gen AI Security Project, "LLM06:2025 Excessive Agency", accessed 16 November 2026: https://genai.owasp.org/llmrisk/llm062025-excessive-agency/
  • OWASP Gen AI Security Project, "LLM01:2025 Prompt Injection", accessed 16 November 2026: https://genai.owasp.org/llmrisk/llm01-prompt-injection/
  • Model Context Protocol, "Authorization", specification 2025-06-18, accessed 16 November 2026: https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization
  • IETF, RFC 8693, "OAuth 2.0 Token Exchange", January 2020, accessed 16 November 2026: https://www.rfc-editor.org/rfc/rfc8693.html

Working on something in this space?

ITMTB works with enterprises on AI, cloud, security, and digital transformation — from strategy through deployment.

Explore More Insights

AI Agents and Enterprise Cybersecurity: Risks, Governance, and Operational Controls

AI Agents and Enterprise Cybersecurity: Risks, Governance, and Operational Controls

Read More
MCP for Business: Can AI Agents Become a New Software Distribution Channel?

MCP for Business: Can AI Agents Become a New Software Distribution Channel?

Read More
Catalogue Visibility for AI Agents: Is Your Catalogue Findable When the Searcher Is an Agent?

Catalogue Visibility for AI Agents: Is Your Catalogue Findable When the Searcher Is an Agent?

Read More
Cybersecurity Risk Assessment for Enterprises: Framework, Operational Risks, and Implementation Guide

Cybersecurity Risk Assessment for Enterprises: Framework, Operational Risks, and Implementation Guide

Read More