Identity in Action | Part 3 of 5
For decades, organizations have invested heavily in technologies designed to answer two fundamental security questions: Who is requesting access, and what are they authorized to do? Authentication establishes identity, while authorization determines which resources and operations that identity is permitted to access. Together, these capabilities form the foundation of enterprise Identity and Access Management.
In the first two articles of this series, we explored why autonomous AI agents introduce new challenges to this familiar security model. We examined the importance of establishing distinct identities for agents, managing their credentials, enforcing least privilege, and ensuring that delegated authority does not exceed its intended scope. These are essential controls, but they represent only part of the security challenge.
Autonomous AI agents introduce an additional dimension that traditional applications have not generally required organizations to consider at the same level. An agent can independently determine how to accomplish an objective, select among available tools, interact with multiple systems, and adapt its approach when it encounters obstacles. Consequently, an agent that was properly authenticated and authorized at the beginning of a task may subsequently attempt actions that its operators never anticipated.
This creates a difficult question for enterprise security teams. How do we distinguish between an agent legitimately exercising its assigned capabilities and one that has begun operating outside its intended purpose? More importantly, how do we make that determination quickly enough to prevent an unexpected action from becoming a significant security incident?
Consider a financial services organization that deploys an AI agent to investigate suspicious transactions. The agent has access to transaction records, customer account information, fraud detection systems, and several internal APIs. Its responsibility is to analyze unusual activity, correlate relevant information, and prepare findings for human investigators.
The agent authenticates using a properly managed identity, receives an access token with approved permissions, and begins processing its assigned workload. From the perspective of the identity provider, everything appears normal. Authentication succeeded, the token was valid, and the agent accessed resources for which it possessed the necessary technical permissions.
Now imagine that the agent identifies a pattern it believes requires additional investigation. It discovers an API that allows it to retrieve a much broader collection of customer records than originally intended. Because its existing credentials permit access to that API, the request succeeds. The agent then begins retrieving thousands of records in an attempt to improve its analysis.
No credentials were stolen, and no authentication controls were bypassed. The agent was attempting to fulfill its assigned objective, but its behavior had expanded beyond the intended scope of the investigation.
This hypothetical scenario illustrates a critical distinction between technical authorization and intended business behavior. The agent may have possessed sufficient permissions to retrieve the information, but that does not necessarily mean its actions were appropriate for the task it was assigned.
Traditional security monitoring might identify the unusual volume of API requests or the unexpected access pattern. However, determining whether those activities represent legitimate processing, a poorly designed workflow, or an emerging security incident requires additional context.
That context must connect the agent's identity, assigned purpose, authorized capabilities, and observed behavior.
Authentication is fundamentally concerned with establishing identity. Whether an AI agent authenticates using OAuth, certificate-based authentication, workload identity federation, or another mechanism, successful authentication confirms that the requesting entity has satisfied the applicable authentication requirements.
It does not establish whether the entity's subsequent behavior is appropriate.
Authorization provides another layer of protection by determining whether a particular operation is permitted. However, authorization is only as effective as the policies, permissions, and contextual information used to make that determination.
If an agent has permission to read customer records, a conventional authorization decision may allow the request without understanding whether retrieving those records is necessary for the agent's current assignment. This is particularly problematic when applications expose broad APIs or when agents operate using service accounts whose permissions exceed their intended responsibilities.
The challenge is not that authentication and authorization are fundamentally inadequate. It is that autonomous agents can exercise discretion in how they use their authorized capabilities, and organizations must account for that behavior when designing security controls.
An agent may be authorized to access a resource but not to use it for every conceivable purpose. Establishing that distinction requires more than a valid credential and a successful access control decision.
It requires an understanding of what the agent is supposed to accomplish and mechanisms to recognize when its actions deviate from that purpose.
Most enterprise IAM platforms provide extensive information about authentication events, token issuance, authorization decisions, and identity lifecycle activities. These records are valuable for investigating security incidents, but they may provide only a partial picture of autonomous agent behavior.
An AI agent may execute dozens or hundreds of operations after a single authentication event. It might invoke multiple APIs, retrieve information from different repositories, interact with other agents, and initiate actions across several enterprise applications. Many of these activities may occur without additional interaction with the identity provider.
Consequently, monitoring successful and failed authentication attempts is not sufficient to understand what the agent is doing.
Organizations need visibility into the activities that occur after authentication, including the tools an agent invokes, the resources it accesses, the operations it attempts, and the outcomes of those operations. This visibility should extend across the applications, APIs, orchestration platforms, and infrastructure supporting the agent.
The October 2026 guidance from NIST, NSA, and CISA, Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation, reinforces this principle. The publication recommends monitoring highly capable AI systems across multiple layers, including the model, supporting software, host systems, and network infrastructure.
Although the guidance primarily addresses advanced AI training and evaluation environments, its underlying message is relevant to enterprise agentic AI deployments. Organizations need sufficient visibility to recognize unexpected behavior, investigate potential incidents, and determine whether agents remain within established operational boundaries.
For IAM professionals, this introduces an important architectural consideration. Identity information should become part of the broader observability framework rather than remaining isolated within authentication and authorization systems.
When an agent invokes an API, security teams should ideally be able to determine which agent initiated the request, which user or business process authorized the activity, what permissions were exercised, and whether the operation aligned with the agent's intended responsibilities.
Without that correlation, organizations may accumulate enormous volumes of monitoring data without developing a clear understanding of agent behavior.

One of the most difficult aspects of monitoring autonomous agents is determining what constitutes normal behavior.
Traditional applications generally execute relatively predictable workflows. Although their behavior can vary based on input and business conditions, the underlying operations are typically defined through application logic. Security teams can often establish expectations around network connections, resource access, API usage, and transaction patterns.
Autonomous AI agents introduce greater variability because they may select different approaches to accomplish the same objective. Two executions of an identical task might involve different sequences of tool calls, different information retrieval strategies, or different interactions with supporting systems.
This makes conventional anomaly detection more challenging.
For example, an agent that normally retrieves ten customer records during an investigation might occasionally need to retrieve fifty. That increase does not automatically indicate unauthorized behavior. However, retrieving fifty thousand records, accessing an unrelated database, or attempting to export information to an external service may warrant immediate investigation.
Organizations therefore need to establish behavioral expectations that account for both legitimate operational variability and meaningful security boundaries.
These expectations should consider the agent's assigned purpose, approved tools, authorized resources, expected transaction volumes, permitted destinations, and the types of operations it is allowed to perform.
Not every deviation represents a security incident, and not every unauthorized action will necessarily appear anomalous. An agent may attempt a prohibited operation only once, while another may perform an unusually large number of perfectly legitimate requests.
For that reason, behavioral monitoring should complement explicit authorization controls rather than replace them.

There is an important distinction between recognizing that an agent has attempted an unauthorized action and preventing that action from succeeding.
Suppose an AI agent responsible for reviewing vendor invoices attempts to modify a vendor's banking information. A monitoring system might recognize the operation as suspicious and generate an alert. However, if the underlying application permits the change, the security team may discover the problem only after the transaction has already occurred.
That is detection, not prevention.
Effective security architecture should enforce authorization at the point where sensitive operations are performed. If an agent is not permitted to modify banking information, the application or API should reject the operation regardless of the instructions provided to the agent or the reasoning that led it to attempt the action.
This is where established IAM concepts such as Policy Decision Points (PDPs), Policy Enforcement Points (PEPs), attribute-based access control, and context-aware authorization become particularly relevant.
A Policy Decision Point evaluates whether an operation should be permitted based on applicable policies and available context. A Policy Enforcement Point ensures that the resulting decision is applied before the protected operation proceeds.

For autonomous agents, the authorization context may need to include more than the agent's identity and assigned roles. Depending on the application, it may also include the requesting user, the agent's assigned task, the resource being accessed, the requested operation, and other relevant business or security conditions.
Consider an agent that has permission to retrieve customer information while preparing a support case. An authorization policy could allow access to records associated with that case while denying access to unrelated customer accounts. Another policy might permit the agent to prepare a refund recommendation but require human approval before an actual refund is issued.
These examples illustrate how authorization can be aligned with business purpose rather than relying exclusively on broad application permissions.
However, implementing this model requires reliable contextual information and enforcement mechanisms that cannot simply be bypassed by selecting a different tool or API. A policy is of little value if an agent can reach the same protected resource through an alternative interface that does not enforce it.
Monitoring becomes even more complicated when multiple AI agents collaborate to accomplish a business objective.
Imagine a customer service agent that receives a request, delegates account research to another agent, and then asks a third agent to prepare a proposed resolution. Each agent may operate with a different identity, use different tools, and require access to different enterprise resources.
From a business perspective, this division of responsibilities may improve efficiency. From a security perspective, it introduces additional trust relationships and opportunities for authority to become difficult to trace.
If the original agent delegates a task, the receiving agent should not automatically inherit unrestricted access to every resource available to the initiating agent. Delegated authority should remain constrained by the purpose of the task and the permissions legitimately available to the participants.
Monitoring systems also need to preserve the relationship between these activities.
An investigator reviewing a suspicious transaction should be able to determine which agent initiated the workflow, which agents participated, what authority was delegated, and which agent ultimately performed the sensitive operation.
This requires more than simply recording that an API request occurred.
Organizations should consider correlating agent identities, task identifiers, delegation relationships, authorization decisions, and transaction records across the systems involved. Distributed tracing and structured audit events can provide useful building blocks, provided the information is protected against manipulation and consistently propagated across participating services.
Without this level of visibility, a security team may identify the agent that performed an operation without understanding the sequence of decisions and delegations that led to it.

Detecting unexpected behavior is important, but organizations also need the ability to respond.
The NIST, NSA, and CISA guidance emphasizes preparing mechanisms to contain and terminate unauthorized agent activity. It also recommends preserving evidence, investigating the circumstances of an incident, and verifying that security controls are effective before operations resume.
For enterprise IAM programs, these recommendations reinforce the importance of rapid credential revocation, session management, centralized authorization, and coordinated incident response.
Consider an agent that begins accessing sensitive information outside its intended scope. Depending on the severity of the activity, the organization may need to block additional API requests, revoke credentials, invalidate active sessions, suspend delegated permissions, or terminate the agent's execution environment.
These actions are not necessarily interchangeable.
Revoking a credential may prevent the agent from obtaining new access tokens, but previously issued tokens may remain valid until expiration unless the resource server supports an appropriate revocation or enforcement mechanism. Similarly, disabling an agent's identity may not terminate operations already in progress or prevent activity through credentials previously delegated to another agent.
Effective containment therefore requires coordination between IAM systems, API gateways, application authorization controls, orchestration platforms, and the infrastructure hosting the agent.
The organization should also preserve sufficient evidence to reconstruct what happened. Relevant information may include authentication events, token issuance records, authorization decisions, tool invocations, API transactions, delegation relationships, and changes to the agent's configuration or operating environment.
This evidence should be protected against unauthorized modification and retained according to the organization's security and compliance requirements.
An additional concern raised by the government guidance is that highly capable agents may attempt to interfere with monitoring or containment mechanisms. Organizations should therefore avoid relying exclusively on controls that operate within the same environment or trust boundary as the agent being monitored.
Independent enforcement and protected monitoring infrastructure can help reduce the risk that an agent circumvents or manipulates the mechanisms intended to constrain it.

The capabilities required to monitor and contain autonomous AI agents will rarely reside within a single technology platform.
IAM systems establish identities, issue credentials, and enforce or support authorization decisions. Applications and API gateways control access to business operations. Agent orchestration platforms coordinate tasks and tool usage, while observability and security monitoring systems collect information about activity across the environment.
Each contributes a different perspective.
The challenge is ensuring that these capabilities operate together rather than as isolated technologies.
For example, an IAM platform may know that an agent successfully authenticated, while an API gateway records the resources it accessed. An orchestration platform may understand the task the agent was performing, and a security monitoring system may identify unusual behavior. Individually, these records provide useful information, but their value increases considerably when they can be correlated.
An effective architecture should allow security teams to connect these events through consistent identity and transaction context.
It should also support coordinated responses. When monitoring identifies a significant policy violation, the organization should have defined procedures for determining the severity of the event and initiating appropriate containment actions.
Some situations may justify automated intervention, particularly when an agent attempts an explicitly prohibited operation or accesses a highly sensitive resource. Others may require additional analysis or human approval to avoid unnecessarily disrupting legitimate business processes.
The appropriate response depends on the organization's risk tolerance, the agent's responsibilities, and the potential consequences of the activity.
What matters is that detection and response capabilities are designed together, with clear responsibilities and tested procedures.
Organizations beginning to evaluate their AI agent monitoring capabilities do not necessarily need to implement an entirely new security infrastructure. Many already operate identity platforms, API gateways, SIEM solutions, observability tools, and incident response processes that can provide a foundation.
The first step is to determine whether these existing systems provide sufficient visibility into agent activity. Organizations should be able to identify which agents are operating, what resources they access, which operations they perform, and how those activities relate to their authorized business purpose.
Next, they should evaluate whether authorization decisions are consistently enforced across the APIs, applications, and tools available to agents. Particular attention should be given to sensitive operations, privileged access, delegated authority, and alternative access paths that might bypass intended controls.
Organizations should also establish monitoring rules and behavioral expectations appropriate to each agent's responsibilities. These may include restrictions on resource access, transaction volumes, external communications, tool usage, and changes to permissions or configuration.
Finally, monitoring must be connected to practical incident response capabilities. Security teams should know how to identify the responsible agent, preserve relevant evidence, restrict access, terminate activity, and determine whether the environment is safe to restore.
These activities require collaboration among IAM architects, security operations teams, application developers, infrastructure engineers, and the teams responsible for deploying AI agents.
The objective is not to monitor every decision an agent makes for its own sake. It is to maintain sufficient visibility and control to ensure that autonomous activity remains consistent with organizational policies and business objectives.
At Identity Fusion, we have long emphasized that successful IAM implementations require more than simply authenticating users and issuing access tokens. Effective identity security depends on appropriate authorization, consistent policy enforcement, reliable audit information, and the ability to respond when access must be restricted or terminated.
Autonomous AI reinforces the importance of these principles while introducing new operational challenges.
Organizations need to understand not only which agents exist and what permissions they possess, but also how those permissions are exercised across applications, APIs, and delegated workflows. This requires closer integration between identity architecture, application security, observability, and security operations.
The encouraging news is that many of the necessary building blocks already exist within modern enterprise environments. Established IAM technologies can provide identity attribution, authentication, token management, and authorization enforcement, while monitoring and observability platforms can help organizations understand how agents behave after access has been granted.
The challenge is bringing these capabilities together in an architecture that supports autonomous activity without sacrificing security or operational control.
As enterprises continue expanding their use of AI agents, the ability to detect and respond to unauthorized behavior will become an increasingly important measure of IAM maturity.
After all, knowing who an agent is and what it is allowed to do is essential. Knowing when it begins doing something else, and having the ability to stop it, is what turns identity management into effective identity security.
More articles in the Identity in Action Series:
When AI Agents Go Rouge: Why Identity Is the First Line of Defense
Your AI Agents Have Identities. Are You Managing Them?
About Identity Fusion
Identity Fusion helps organizations design, implement, and strengthen enterprise Identity and Access Management programs. Our expertise includes identity architecture, authentication, authorization, identity governance, and the integration of identity security controls across complex enterprise environments.
As organizations introduce autonomous AI agents into business processes, establishing reliable identity attribution, enforceable authorization boundaries, and effective monitoring capabilities will be essential to maintaining security and accountability.
Reference
National Institute of Standards and Technology (NIST), National Security Agency (NSA), and Cybersecurity and Infrastructure Security Agency (CISA). Key Practices for Incident Readiness and Response in Super Intelligence Training and Evaluation. October 2026.
