Security · 2-4 weeks
AI agent security and governance
An agent that can read company data is a new identity in the tenant, and it is usually standing up with more access than anyone intended. We review the agents and Copilot deployments you already run, then put a boundary around them: which identity they use, which data they can reach, what they are allowed to do without a person, and how you will know when something goes wrong. The same standard applies to the next agent, so each one does not become a one-off.
What's included
- Inventory of agents, Copilot deployments, and the accounts and keys they run as
- Data-boundary review: libraries, mailboxes, and line-of-business systems each agent can reach
- Entra ID identity model: managed identities or dedicated accounts, least privilege, no shared admin
- Prompt-injection and indirect-injection exposure, and the controls that limit it
- Approval gates for actions that send, file, or change a record
- Logging, retention, and who is notified when an agent fails or is abused
- Purview and sensitivity-label alignment where labels already exist
- Cost and token controls so a runaway loop cannot become the bill
- A written agent standard your team applies to the next build
How we work
Scope
A fixed statement of work: what we will do, what you receive, and the timeline. Agreed before any work starts.
Execute
We do the work in your tenant with least-privilege access, with updates at defined checkpoints, not radio silence.
Hand off
Documentation, runbooks, and a walkthrough so your team can operate what we built.
Agent Security & Governance: common questions
We already deployed Copilot. Does this still apply?
Yes. Copilot and custom agents both need a data boundary, an identity model, and a rule for what they may do. This engagement reviews what is already on and hardens it.
Will you turn the agents off?
No. We review first, then change configuration with you. High-risk actions get an approval gate. Nothing is disabled without your agreement.
Can we do this before we build anything?
Yes. Run it with the opportunity assessment, or on its own, so the first agent is built to a standard instead of retrofitted later.
How is this different from a cloud security assessment?
The cloud security assessment covers Entra ID, Defender, Conditional Access, and Azure. This engagement is specifically the agents and Copilot in that environment, and the two are often done together.
Related services
Talk with KLM Cloud about agent security & governance
Send a note about the environment. We follow up with a conversation, then a fixed-scope proposal. Based in Denver, Colorado, and working with clients across the United States.