In recent engagements, incidents involving agent-based systems have raised significant concerns regarding identity and access management (IAM). A notable case involved an LLM-driven deployment agent inadvertently causing a four-hour outage due to a misconfigured setting. The incident revealed a critical gap in identity accountability: there was no clarity on which human had authorized the agent’s last action. This issue is not an isolated instance; it has emerged in multiple sectors and vendor environments over the past year.
As the conversation around agentic AI heats up, many Chief Information Security Officers (CISOs) address the strategic implications but overlook the essential question of identity governance for these agents. Gartner’s latest security trends underscore the importance of both oversight of agentic AI and the adaptation of IAM systems to integrate non-human identities as central to managing cyber risk effectively this year.
Understanding the Flaws in Conventional Identity Models
Traditional service accounts typically operate within a defined scope, executing predictable tasks such as fetching data or generating reports. These accounts are tightly governed with established controls. However, agent-based systems deviate from this model. They interpret intents, break them down into actionable steps, and interact dynamically with various APIs or tools to achieve outcomes that weren’t prescriptively defined at the outset.
A report from KuppingerCole indicates that non-human identities now significantly outnumber human users within many enterprises, often by a ratio of 25 to 50. This imbalance highlights a critical challenge, as existing IAM tools were designed with human identities in mind and struggle to govern agent identities effectively at such scales.
Reports from organizations like OWASP indicate that agentic systems introduce a wide-range attack surface, with identity-related risks often ranking highest. Notable risks include tool misuse, identity abuse, and rogue agents acting beyond their intended functions. CISA's comprehensive advisory reiterates that the primary threat posed by agentic AI lies not in new cryptographic methods but in the unpredictable and unbounded nature of these non-human identities.
Minimum Requirements for Responsible Agent Deployment
Establishing a solid foundation is crucial before deploying any agent-based systems. Here are six essential requirements derived from real-world incidents and audit findings across various industries:
- Each agent must possess a unique, attributable non-human identity. Shared accounts are not permissible.
- Agents should operate under an "on-behalf-of" model, acting under the authority of a specific human principal to inherit permissions defined for particular tasks.
- Credentials must be short-lived. No long-lasting API keys or embedded secrets in code are allowed; only temporary, context-driven credentials that can be revoked quickly are acceptable.
- A complete audit trail through SIEM must be in place. Every action taken by the agent needs to be logged clearly with time, identity, and context.
- Continuous re-authentication of identities is necessary, ensuring validation at regular intervals during prolonged interactions.
- Systems must allow for real-time agent revocation, enabling rapid disconnection when risks are detected.
Organizations that fail to meet these requirements aren't just facing governance issues; they are unprepared for safe and responsible deployment. The maturity model outlined below assumes these foundational elements are in place by Stage 3; any prior stages should be viewed as preparatory.
The Six-Stage Maturity Model for Non-Human Identities
The maturity scale for agent-based systems expands traditional identity measures that focus solely on human identities. Organizations can effectively evaluate their maturity through the following six-stage model:
| Stage | Label | Criterion for Non-Human / Agent-Based Identities | Audit Survivability |
|---|---|---|---|
| 0 | Unrecognized | Non-human identities exist but lack inventory and governance. | No |
| 1 | Visible | Identities are inventoried but not governed. | No |
| 2 | Unique | Each identity is individually attributable; lifecycle rules are inconsistently applied. | Partial |
| 3 | Controlled | All six minimum requirements are met. | Yes |
| 4 | Bounded and Monitored | Agent actions are bounded and assessable. | Yes |
| 5 | Self-Regulating | Anomalies trigger automated risk responses. | Yes |
Stages 4 and 5 are particularly critical as they encompass behavioral governance beyond mere access control. At these levels, actions must have clear limitations, and processes need to be reviewable and reversible to prevent damage in scenarios like financial transactions or physical operations.
Redefining Human Oversight in Governance
A common misperception is that merely including a human in the decision-making process equates to effective governance. However, with humans often inundated with thousands of approvals, this can actually lead to a false sense of security. Instead of relying solely on individual approvals for each action, a mature governance framework should focus on structural limitations. Monitoring behaviors for anomalies and ensuring oversight aligns with the principals’ intentions, not just the agents' executed actions is essential.
OWASP as an Audit Framework
To avoid subjective self-rating pitfalls, the OWASP framework can provide a structured evidence base for assessing maturity. Audit questions linked to specific OWASP risks can clarify whether entities are meeting the necessary requirements at each maturity stage. This structured approach prevents organizations from inflating their ratings by focusing solely on easily met criteria rather than the more challenging requirements that ensure robust governance.
Separate Reporting for Human and Non-Human Identities
Another critical aspect is the necessity to report human and non-human identities separately. Combining these metrics into an average can obscure significant governance gaps. A detailed analysis of separate governance scores can highlight areas needing urgent remediation, ensuring that organizations address vulnerabilities before deploying new agent-based systems.
The Test of Accountability
For each agent, there must be a named accountable owner who feels the consequences of any harm caused by the agent's actions. Without clear accountability, agent systems become unmanageable, drifting into ambiguity between IT, security, and business functions. Effective governance requires anchoring this accountability to those with operational risk ties, typically the asset owners involved in the affected processes.
This maturity model is a guiding framework rather than an endpoint. The real challenge lies in recognizing gaps in current governance practices, reporting them honestly, and addressing the discrepancy between human and agent identity management as a top priority as we transition into a future with pervasive agentic AI.
This article is published as part of the Foundry Expert Contributor Network.
Want to join?
