78% of companies have no policy for their AI agents. By 2028, the average Fortune 500 will run 150,000 of them.

Non-human identities (NHIs) are the credentials that let software act on behalf of humans and systems: API tokens, service accounts, autonomous AI agents, scheduled jobs, and the keys they all carry. By 2028, Gartner projects the average Fortune 500 enterprise will run more than 150,000 AI agents. Each of those agents is its own identity, with its own permissions, and almost no one inside the average company is responsible for turning it off.

That is the finding inside a Cloud Security Alliance survey: 78% of organizations lack a documented policy for creating or removing AI identities, and 79% report low or moderate confidence in their ability to stop an attack that uses those identities. A separate CSA white paper calls the situation a governance and ownership crisis.

The problem is that the identities are structurally ungovernable in the systems they have today.

Each NHI is born when a developer ships a piece of code that needs to call another service. The developer creates a token, the token gets permissions, the code goes to production, and the developer moves on. There is no step where anyone decides who owns the token, what it is allowed to do, or when it should expire. Aleksandr Krasnov of Ducker Tech Consulting found that a single developer can hold 244 such identities. Multiply that across an engineering organization, and the number reaches six figures before a single AI agent is added.

They become ghost credentials, the term Krasnov uses for dormant NHIs that still hold permissions in the cloud but have no owner, no purpose, and no lifecycle (BackBox). An attacker who finds one has a real key to a real system, with no human to notice the alarm. From there, lateral movement across services and privilege escalation are standard playbook steps.

The traditional fix is identity and access management (IAM), the set of tools that govern human logins. NHIs do not fit that mold. A human user has a manager, a joiner-mover-leaver process, and a single sign-on policy. A token has none of those. It is created by code, owned by no one, and removed only when the cloud bill shows an unused secret, which it usually does not.

Okta agreed to acquire Permiso, a cloud-native identity platform that covers human, non-human, and agentic identities, for about $200 million, per TechCrunch's source. A day earlier, Cyera said it would buy Oasis Security for roughly $1 billion to safeguard the same category of customer. Both deals are framed as consolidation around agentic identity.

Read together with the CSA data, the M&A wave looks more like a market admitting the gap than closing it. Buyers are paying to bolt NHI governance onto existing IAM and data-security platforms, not to replace those platforms. The result, if the trend continues, will be a category defined by the vendors' defaults rather than by the customer's policy. Whoever owns the consolidation wins the question of which controls become the floor.

Krasnov's response has been to give some of the tooling away. His open-source project, NHI Hound, inventories human and non-human identities, maps the trust links between them, and simulates how an attacker could move from one account to another (BackBox). Early pilots with small and mid-sized companies have shown the inventory is what the operators are missing, not a magic detection rule.

The closing window is the 18 to 24 months before 2028. By then, the average Fortune 500 will run 150,000 agents. The companies that have not assigned owners, written lifecycle policies, and audited the dormant credentials in their clouds by then will inherit the defaults that the consolidating vendors ship. Those defaults are better than nothing. They are not the policy the CSA found 78% of organizations do not have.