Recent incidents in enterprises have highlighted a critical gap in identity governance for agentic AI systems. A notable case involved an LLM-based deployment agent that inadvertently caused a four-hour outage through a misconfigured setting. During the incident review, the question of which human authorized the agent's actions proved difficult to answer, signaling a broader trend observed across various sectors. Security leaders are increasingly aware of the potential risks posed by agentic AI, but many still overlook the foundational elements of identity management for these non-human entities.
A recent report by Gartner outlines essential trends in cybersecurity, emphasizing the need for oversight of agentic AI and adaptive identity and access management (IAM) strategies. This six-stage maturity model for non-human identities (NHIs) proposes a structured approach to understanding and managing these entities effectively, establishing minimum requirements before deploying any agent-based system into production.
The Challenge of Agent-Based Systems
Traditional service accounts are relatively simple, performing limited, predefined tasks effectively governed by established controls like routine credential rotations and audits. However, the complexity of agent-based systems introduces new challenges. These systems operate autonomously by interpreting commands and taking actions based on their assessments rather than following strict, predefined steps. According to KuppingerCole’s recent analysis, NHIs have surpassed the number of human users in many organizations, presenting significant governance difficulties that existing IAM models weren’t designed to handle.
The OWASP GenAI Security Project has identified key vulnerabilities emerging from the misuse of identities within these systems. With threats including identity and privilege abuse, recurrent issues surrounding rogue agents, and inherent risks associated with agent-based supply chains, the need for a robust identity governance framework becomes apparent. Recent advisories from cybersecurity authorities reinforce that privilege risk is foundational in these scenarios, requiring a re-evaluation of our IAM strategies to address the unique challenges posed by agentic AI.
Necessities Before Deployment
Organizations must adhere to six minimum requirements for governance to ensure a responsible deployment of agent-based systems:
- Each agent must possess a unique, attributable identity; shared accounts are not permissible.
- Permissions should follow an on-behalf-of model, allowing agents to act under a named principal’s authority without independent standing.
- Long-lived credentials are banned; instead, short-lived, context-bound certificates must be used.
- An audit trail through Security Information and Event Management (SIEM) integration is essential for accountability.
- Continuous re-authentication mechanisms should be in place for long-running agents based on risk assessments.
- Real-time revocation capabilities are crucial to mitigate potential incidents quickly.
Achieving compliance with these requirements represents the baseline for deploying agent-based systems in a secure environment. Any organization failing to meet these standards has a fundamental readiness issue rather than merely a governance problem.
Understanding the Six-Stage Maturity Model
Most maturity assessments conveniently focus on human identities, often neglecting the complexities introduced by NHIs. Organizations might score high in one dimension while remaining blind to deficiencies in the other. This new model proposes a cumulative six-stage framework designed to address these distinct identity types effectively, ensuring comprehensive governance practices.
| Stage | Label | Criteria for Non-Human Identities | Audit Readiness |
|---|---|---|---|
| 0 | Unrecognized | Non-human identities exist without inventory; long-lived keys, no accountability. | No visibility into activities. |
| 1 | Visible | Identities are recognized but lack independent governance. | No accountability per agent. |
| 2 | Unique | Every identity stands alone with lifecycle rules inconsistently enforced. | Partial accountability. |
| 3 | Controlled | All six minimum requirements are satisfied, establishing a defensible posture. | Yes — meets basic audit standards. |
| 4 | Bounded and Monitored | Agent actions are limited, and every step can be tracked and reviewed. | Audit-ready documentation. |
| 5 | Self-Regulating | Automatically detects anomalies, requiring an accountable owner for each agent. | Meets the highest standard. |
Stages 4 and 5 highlight the evolution of governance — limiting the scope of agent actions, ensuring accountability, and enforcing behavior monitoring capabilities. A focus on structural constraints over manual approvals is necessary for scaling governance effectively. This model urges organizations not merely to perform checks but to embed discipline and best practices in their operational protocols.
The Importance of Governance Beyond Human Oversight
A common misconception is that having a human involved in the decision-making loop guarantees proper governance. However, if a human is overwhelmed by the volume of agent actions requiring approval, it effectively turns into approval automation rather than genuine oversight. Mature governance structures emphasize defining clear constraints around what agents can do and monitoring their behavior, ensuring accountability remains tied to the principles set by human oversight.
Leveraging OWASP for Audit Readiness
To avoid subjective self-assessments during maturity evaluations, organizations can translate OWASP's findings into actionable audit questions, anchoring each maturity stage with evidence-based metrics. The criteria established can help organizations identify gaps, ensuring they adhere to rigorous standards while progressing through the maturity model.
Separating Human and Non-Human Identity Reporting
One critical reporting decision revolves around the need to avoid aggregating scores between human and non-human identity governance. A clear distinction allows organizations to pinpoint areas needing attention, avoiding an inflated sense of security that may arise from combining diverse governance stages into a single average score. This level of clarity emphasizes the significance of addressing weaknesses in agent governance, which represent some of the most unpredictable risks in identity management.
Identifying Accountability for Agents
A vital question to address during engagements is: who is accountable for each deployed agent? Without a designated accountable owner, systems risk becoming orphaned assets, complicating incident responses when issues arise. Establishing accountability ensures operational risks are owned by the relevant personnel within business functions, reducing ambiguity surrounding roles and responsibilities.
The proposed maturity model lays a foundation for organizations looking to enhance their governance of agent-based systems. The primary starting point should be honest assessment and reporting of identity management practices, encouraging a proactive approach to security challenges presented by agentic AI.
