Enterprise security teams are currently rushing to put digital badges on AI agents, but they are doing so without the security cameras needed to watch how those agents behave. This push to secure the identity layer follows recent industry momentum around agent risk management and the October governance sprint, but it reveals a structural imbalance: we are effectively giving these agents credentials before we have the tools to validate their actions.
In just a few months, three major enterprise vendors have rolled out distinct approaches to agent identity. Microsoft led the charge in May 2026 by rolling out Microsoft Entra Agent ID, which automatically provisions identities for agents created in Copilot Studio. This was followed by Okta Agent SSO, which launched on September 22, 2026, at Oktane 2026, moving away from static API keys toward short-lived, identity-governed tokens. Most recently, ServiceNow AI Passport was announced on October 15, 2026, introducing a cryptographically signed credential that agents consult at runtime to verify their authorized scope.
The data suggests a profound disconnect between how executives perceive their security posture and the reality on the ground. According to the Okta AI Agents at Work 2026 survey, 96% of executives express confidence in their identity and access management for non-human identities. Yet, only 34% of organizations actually apply the same security controls to their digital labor force as they do to their human employees.
This gap is even more stark when looking at visibility. The Saviynt CISO AI Risk Report 2026 found that 92% of CISOs lack full visibility into AI agent identities. When you consider that 71% of these organizations have AI systems accessing core systems like ERP or CRM, the lack of governance is striking. Only 16% of those organizations govern that access effectively, and 86% do not enforce access policies for AI identities at all.
There is a growing consensus in the industry, as discussed at the Identiverse 2026 conference, that the market is currently conflating visibility with enforcement. Knowing an agent exists is not the same as controlling its actions. The current human-centric IAM models are proving insufficient for non-deterministic AI agents, and the measurement layer — the ability to test, evaluate, and validate agent behavior — remains critically underdeveloped. In fact, the NIST AI Risk Management Framework MEASURE function has been scored as critical, with a lack of processes for drift or behavioral tracking.
This raises difficult questions for the enterprise. Who changes their behavior when these identity layers are deployed? Who gains slack in their daily operations, and who loses leverage? Most importantly, who carries the cost when an agent, now properly identified but poorly governed, drifts into unauthorized territory?
This analysis builds on our prior coverage of agent risk management as a funded category and the October governance sprint, but it focuses specifically on the identity layer as the foundation. However, the identity layer is not the house. As we move forward, the pressure will shift toward the validation layer. Organizations will need to move beyond simply issuing passports or tokens to agents and start implementing rigorous behavioral tracking. A practical caveat: even with perfect identity, the non-deterministic nature of AI means that traditional static policies will likely struggle to keep pace. The next phase of the agent economy will be defined by those who can bridge the gap between knowing who an agent is and knowing exactly what it is doing in real-time.
