AI Agents and Non-Human Identities: Governance Challenges

Cibersegurança · 2 min

AI agents are beginning to execute tasks with limited autonomy. The critical question is how to control identity, permissions, and oversight without hindering innovation.

AI Agents and Non-Human Identities: Governance Challenges

TL;DR

AI agents introduce new operational identities into organizations.
Governance must cover identity, permissions, and traceability.
The principle of least privilege remains essential.
Human oversight should be adjusted to the risk of each action.
Revocation and auditing are fundamental to maintaining control.

The new question: who authorized the agent?

For years, artificial intelligence applications were mostly limited to analyzing information or supporting human decisions. With the emergence of AI agents, this reality is beginning to change: these systems can execute actions on business applications, interact with other services, and trigger processes autonomously within defined limits. This evolution introduces a new challenge: __how to govern identities that no longer correspond to human users__.

Non-human identities require their own control

An AI agent should not be treated as a mere extension of the user invoking it. It must have *its own identity*, defined scope, proportional permissions, and independent logs. This distinction is important because the risk is not limited to the supply chain of models, libraries, or vendors. Operational risk arises when the agent acts in real systems with credentials, integrations, and execution capability.

Least privilege also applies to agents

In practice, this means assigning only the __strictly necessary permissions__ for each task, prioritizing temporary credentials whenever possible, and avoiding shared accounts or permanent privileges. Wherever supported by the platform, managed identity mechanisms or short-lived credentials reduce the exposure associated with persistent secrets.

Human oversight should not be generic

Requiring human approval for all actions can make the agent irrelevant. Not requiring approval for any action can create unnecessary exposure. The decision should be based on risk: - *informational actions* can be automated; - *reversible actions* can have limits and logging; - *sensitive actions* must require explicit validation. In distributed environments, where multiple agents interact with applications, APIs, and cloud services, architectural models like the Cybersecurity Mesh can facilitate the consistent application of identity, authentication, and monitoring policies.

Logging, traceability, and revocation

Agent governance must include **clear logs**: which agent acted, on behalf of which user or process, with what data, what tool it invoked, and what the result was. These logs should be useful for auditing, incident response, and continuous improvement. There must also be a simple mechanism to suspend, limit, or revoke an agent when the business context changes, the integration is no longer needed, or anomalous behavior is detected. Without practical revocation, governance becomes merely _documentary_.

Conclusion: autonomy with explicit boundaries

As AI agents take on a more active role in business operations, identity ceases to be merely a technical attribute and becomes a central element of governance. Preparing for this change from the outset can help organizations adopt automation more securely, transparently, and sustainably.