The first security question about an AI agent is often: What can it access? That matters, but it is not the first question organisations should answer.
Before deciding which systems an agent can reach, leaders need to know why it exists, what outcome it is authorised to pursue and who remains accountable for its actions. Without that context, identity and access management (IAM) can become a way to grant permissions without establishing whether they are justified.
Agents are moving beyond assisting employees and beginning to act across business systems. They need their own identities, controlled permissions and a means of revoking their authority. Those technical controls, however, cannot decide whether the agent’s purpose is legitimate or whether its authority remains appropriate.
Organisations should give each agent a distinct, auditable identity, a named human sponsor and permissions tied to an approved outcome. They should review that authority when the agent’s purpose, ownership, capabilities or risk changes. Identity is the technical core of agent security, but business governance defines what that identity should be allowed to do.
Identity Is the Technical Core of Agent Security
Agents are useful because they can act. Unlike an assistant that produces content for a person to assess, an agent may use tools, retrieve organisational data or perform actions across a workflow, depending on its design and permissions. This creates a security problem of delegated authority, not just access.
Microsoft’s Digital Defense Report 2026 recommends giving agents their own identities rather than having them borrow employee credentials. It also recommends making those identities auditable and recording information such as the agent’s creator, purpose and sponsor. This makes it easier to distinguish a person’s actions from an agent’s and gives security teams a clear entity to review, investigate or disable.
A mature IAM programme already covers authentication, authorisation, lifecycle management and access governance. Agent identities extend the need to manage digital identities and control access to systems, applications and data across a new population that must be discovered and governed alongside employees, applications and traditional workloads.
Creating an identity is only the starting point. Permissions should reflect the agent’s approved role, with access limited to the resources and actions needed for that role. Microsoft recommends least privilege, time-limited assignments and approval workflows, alongside removing access that is no longer used. These controls matter because an agent may exercise its permissions differently from a person.
The useful question is not simply what access a workflow technically requires. Leaders should ask what authority the agent legitimately needs to achieve its approved outcome. That distinction makes least privilege a business decision as well as a technical one.
IAM Cannot Decide Why an Agent Should Have Access
IAM can authenticate an entity, enforce access policies and support reviews or revocation. It cannot independently decide whether an agent’s objective is worth pursuing. That decision belongs to the business, with security and technology teams helping translate it into controls.
Bytes' The AI Ambition Gap examines why AI adoption does not automatically translate into organisational productivity. Its analysis focuses on leadership, direction and redesigning work around AI. The security implication is a further question: what authority should an agent receive if the organisation has not defined its purpose, outcome and accountable owner?
Access decisions should begin with the business problem, then define the authorised outcome and accountable owner before identifying permitted actions, resources and permissions. Starting instead with a platform’s capabilities risks making available technology, rather than the intended outcome, the basis for security design.
Consider an agent that prepares material for a monthly financial review. The fact that it can connect to a finance system does not mean it should be able to change financial records. If its approved role is to compile and summarise information, read access to a defined dataset may be enough. Write access would need a separate justification and risk decision.
Define an Identity Contract for Each Agent
Organisations can connect the business decision to technical controls through an identity contract. This is a governance model, not a Microsoft product feature or replacement for IAM. It records the authority delegated to an agent before that authority is translated into permissions.
An identity contract should cover six areas:
- Business purpose: State the problem the agent exists to address. ‘Improve productivity’ is too broad to guide access decisions, while a defined task, such as preparing approved customer-service cases for human review, gives security teams a clearer boundary.
- Authorised outcome: Define what the agent may accomplish and where its authority ends. An agent that retrieves information may need different controls from one that changes records or initiates transactions.
- Human sponsor: Name the person accountable for the agent’s continued suitability. Make technical responsibility clear too, so business ownership of the outcome is distinct from responsibility for configuration, permissions, monitoring and response.
- Identity and permissions: Provision a distinct identity and restrict it to the resources needed for its approved function. An agent that retrieves documents poses a different risk from one that can alter configurations, approve transactions or invoke privileged tools.
- Operating boundaries: Record the systems, applications, data and tools available to the agent. These boundaries help teams understand how the agent could affect the wider environment, particularly when it can interact with other agents.
- Review and revocation: Specify when the contract must be reconsidered, including a change of sponsor, a new integration, an increase in privilege or the end of the business requirement. Microsoft recommends lifecycle controls that test whether an agent is still needed, whether its access remains appropriate and whether it has a valid sponsor.
The contract should remain current as the agent changes. A document completed at deployment will not reflect new tools, workflows or permissions unless someone is responsible for reviewing it.
“Governing AI agents is about more than giving each agent an identity and controlling its permissions. As agents gain greater autonomy, organisations need to be clear about the outcome each agent is authorised to pursue, who remains accountable for its actions and where its authority should end. Identity gives you the technical control point, but effective governance connects that identity to human ownership, business purpose and clear operating boundaries.”
Agent-to-Agent Access Extends the Trust Chain
Treating agents simply as a larger population of user identities misses an architectural difference. An agentic workflow may involve a human, an agent, a sub-agent, a tool, an application and data. Each connection creates a relationship that security teams may need to identify, authorise and monitor.
A sub-agent should not become trusted merely because it sits downstream of an authenticated human interaction. Nor should an agent carry broad, long-lived credentials for every tool it might use. Microsoft’s least-privilege guidance supports scoped access and validation of authorisation through downstream systems. Organisations should apply the same discipline at each point where authority is delegated.
Security leaders need to be able to trace an action back through the chain to an approved purpose and an accountable owner. If that traceability ends partway through the workflow, the organisation has an accountability gap, even if every component is authenticated successfully. Authentication establishes which identity is acting; governance explains why it is allowed to act.
Review Authority Throughout the Agent’s Lifecycle
Provisioning establishes an agent’s initial access, not a permanent policy. Organisations should reassess whether they still serve their approved purpose, whether their sponsor remains accountable and whether new tools or data have changed their potential impact. Periodic reviews should test the business justification as well as the permissions recorded in the identity system.
Revocation needs equal attention. Microsoft recommends testing how access can be removed when an agent or its owner is compromised. Security leaders should measure how quickly they can identify the agent’s reach and stop it from acting, including across connected systems. Disabling an identity may not, by itself, remove every downstream permission or already-issued credential, so the full containment process needs to be tested.
This is a practical measure of resilience. No organisation can assume every undesirable action will be prevented, but it can establish whether it can identify delegated authority, limit its impact and withdraw access when conditions change.
The Goal Is Accountable Agency
The answer to agent risk is not to remove the capabilities that make agents useful. It is to make their authority explicit and accountable.
The NCSC’s Annual Review 2026 says AI is likely to increase the speed, scale and efficiency of established cyber operations, while human oversight remains important. For organisations deploying agents, this reinforces the need to control what systems can act, on whose authority and within which boundaries.
Identity provides the technical foundation. Each agent should be attributable, its permissions proportionate and its access reviewable and revocable. But identity alone cannot determine whether the agent’s purpose is legitimate or whether its authority still serves the organisation.
That requires business ownership alongside technical controls. Identity tells you which agent is acting. Governance must explain why it is allowed to act. CyberOne can help organisations establish the visibility, ownership, lifecycle and access-control foundations for enterprise AI agents, aligned with wider business and security objectives.