Back to blog
#ai#security#business

AI Agent Identity in 2027: Six Vendors, No Standard

Vendors are shipping AI agent identity and registry products while the standard is still a draft. Here is what to put in place in 2027 without waiting for it.

By Rafael Costa6 min readEnglish
Share
AI Agent Identity in 2027: Six Vendors, No Standard

Here is a question most companies cannot answer: how many AI agents are running against your systems right now, and who authorized each one to do what it does? A year ago that was a theoretical worry. Now it is an audit finding. Agents authenticate to your APIs, hold credentials, and act on their own, and in most shops they do it on shared service accounts that make the logs say "the integration user" instead of a name. When something goes wrong, you cannot tell which agent did it, let alone which human is on the hook.

That gap is why agent identity went from a slide in a security deck to a product category in 2026. The uncomfortable part is the timing: vendors are shipping identity and registry products for agents while the open standard that would make them interoperate is still a draft that keeps getting renewed. You have to make decisions now, in a market that has not settled. This post is about what is actually real, what is still moving, and what to put in place in 2027 so you are not securing a fleet you cannot even see.

Why agent identity became a 2027 problem

A human identity is relatively easy. One person, one login, one trail. An AI agent breaks all three assumptions at once.

  • It runs on borrowed credentials. The fastest way to ship an agent is to hand it a service account that can already reach the CRM, the database and the file store. That works until you need to answer who did what, because every agent on that account looks identical in the log.
  • It acts without a human in the loop. The whole point of an autonomous agent is that it decides and acts on its own. So "the user clicked the button" is no longer the record of intent. Something has to stand in for authorization, and a shared secret does not.
  • It multiplies quietly. A team wires up one agent, then three, then a dozen. We wrote about this as agent sprawl: the average enterprise now runs around a dozen agents, and most of them were never registered anywhere. You cannot govern what nobody wrote down.

Put those together and the number everyone quotes starts to make sense: depending on the survey, only something like 11 to 14% of agentic AI pilots reach production. A large share of the rest stall on exactly this, identity, audit and access control, not on whether the model is smart enough.

The registry you don't have yet

Before you can give an agent an identity, you have to know the agent exists. That sounds obvious and it is the step almost everyone skips. The public Model Context Protocol registry has already grown past 10,000 servers, and the protocol still has no standard discovery mechanism, which means the inventory of what is running inside your own walls is a thing you build, not a thing you query.

This is where the new products are aiming first. RSA's Agent ID platform, reaching general availability in mid-November 2026, leads with a discovery module whose job is to inventory agents and MCP servers across your identity, cloud, endpoint and gateway data, then assign each one a human owner, a risk tier and a lifecycle state. Strip the branding and that is the control every governance framework has been asking for: a list of every non-human actor, each tied to a person who answers for it.

Build the inventory before you buy the tool

You do not need a platform to start the registry. A single shared sheet with one row per agent, its purpose, the systems it can touch, the credential it uses and the human who owns it will expose more risk in an afternoon than most teams expect. Half the entries will be agents nobody remembered. That list is also the thing you hand any identity vendor on day one, so the work is never wasted. Discovery first, identity second, and only then automation.

Six vendors shipping, one standard still in draft

The reason to move carefully is that the plumbing is not finished. Through 2026, RSA, Microsoft, Okta, Auth0 and Descope all shipped agent-identity features, each with its own model of what an agent credential looks like and how delegation works. Meanwhile the IETF's draft for agent delegation tokens, ID-JAG, is still an Internet-Draft, on a version that expires in late November 2026 and will be renewed rather than ratified. Workload identity and agent delegation remain, in the standards bodies' own words, under development.

So you are buying into a category where the vendors agree on the problem and disagree on the format. That is survivable, but it changes how you buy. Favour products that let an agent's identity map cleanly onto the identity system you already run, rather than inventing a parallel directory of agents that will be painful to reconcile when the standard does land. And keep the enforcement point under your control.

Don't let agent identity become a second shadow directory

The failure mode here is predictable: each team adopts a different agent-identity tool, and eighteen months later you have four registries that disagree about which agents exist and who owns them. That is shadow AI wearing a security badge. Pick one source of truth for non-human identity now, even an imperfect one, and route every new agent through it, so consolidation later is a migration and not an archaeology project.

What to put in place now

You do not have to wait for the IETF to make real progress. The moves that pay off are the ones that hold regardless of which format wins.

  1. Give every agent its own identity. No more shared service accounts. One agent, one credential, so the audit log names the actor. This is the single change that turns "something touched the invoice" into "this agent did, at this time, on this person's authority."
  2. Scope it to least privilege. An agent that reads orders does not need write access to payroll. Grant the narrow set of permissions the job requires and nothing more, which is also the foundation of sandboxing an agent before it can break something.
  3. Attach a human owner and a risk tier. Every agent gets a named person who answers for it and a tier that decides how much oversight it needs. If you have not filled the AI agent owner role, this is the forcing function to do it.
  4. Meter and audit at one choke point. Route agent traffic through a gateway so identity, access and cost are enforced in one place rather than reinvented per codebase. That is the case we make for an MCP gateway, and it is where your audit trail actually gets written.
  5. Keep a kill switch and watch behaviour. Identity tells you who; observability tells you what they did and when it drifts. You want to revoke one agent's access in seconds without taking down the rest.

None of this is exotic. It is the discipline that already governs human and service accounts, applied to a new kind of actor before the fleet gets too big to map. The foundational version of this argument, scoping and auditing non-human identities, is in our earlier piece on securing non-human identity; what changed in late 2026 is that the tooling to do it at scale finally shipped, standard or no standard.

Start before the standard lands

The companies that will sail through their first serious agent audit in 2027 are not the ones who waited for a clean specification. They are the ones who could produce a list of every agent, its owner, its permissions and its trail, on the day the auditor asked. That list is cheap to start and expensive to reconstruct after the fact.

If you are putting agents into production and want identity, access and audit built in from the first one rather than retrofitted after a scary finding, tell us how your stack is set up and we will help you put the controls where they belong, in a way that survives whatever the standards bodies finally agree on.

#ai#security#business
Share this article
Rafael Costa

Written by

Rafael Costa

Software Engineer & Technical Writer

Rafael is a software engineer at Lusivision who writes about web development, cloud architecture and applied AI. He has spent over a decade shipping production software for companies across Europe and enjoys turning hard technical topics into clear, practical guides.

View all articles

Related articles

Newsletter

Stay in the loop

Occasional notes on software, design and what we're building. No spam — unsubscribe anytime.