Skip to content
Okta vs Entra

Reviewed 10 November 2026

Shadow agent discovery differs across Okta and Microsoft

Reviewed 2026-11-10 · Editorial assessment by the editors

By the editors. Published 2026-11-10. 5 min read. Tags: AI agents, Discovery, Okta, Microsoft Entra.

Okta describes shadow agent discovery through a browser plugin and EDR integrations, including coding agents such as Claude Code and Cursor. Microsoft’s October announcement adds identity patterns for previously unknown agents discovered through security sensors. Neither description should be treated as proof that discovery alone approves or governs every action.

Compare the starting evidence rather than the label

Shadow agent discovery begins with a practical problem: the security team needs a way to learn about agents it did not deliberately register. Okta and Microsoft describe relevant capabilities, but the verified material places them in different contexts. Okta describes endpoint discovery mechanisms. Microsoft describes additional identity patterns that include previously unknown agents discovered through security sensors. Compare those exact statements before treating the two as equivalent inventory products.

For a proposed evaluation, define what counts as a useful finding. It should be possible to explain what was observed, where it was observed and what decision follows. Do not assume a discovery label establishes a complete record of permissions or a named owner. Those are separate pieces of evidence to request. An observation becomes operationally useful when someone can decide what to do with it.

  • Identify the observation source and the environment in which the agent appeared.
  • Request the identity record produced from that observation.
  • Record the person responsible for reviewing the finding and deciding its disposition.

Read Okta’s endpoint mechanisms with their release status

Okta’s September release lists Shadow AI Agent Discovery for Endpoints as planned to be generally available in Q3. The release does not say whether Q3 is fiscal or calendar. Its later NIST checklist describes shadow agent discovery using a browser plugin and EDR integrations. The checklist says this finds coding agents such as Claude Code and Cursor. These details explain the described observation paths without establishing coverage for every endpoint or agent.

Keep the mechanism and the availability statement together in the evaluation record. A browser plugin and an EDR integration are not interchangeable descriptions, so ask which is relevant to the intended environment. Request supported-environment documentation and evidence of what appears when an agent is found. Do not use the existence of an implementation description to silently change the earlier planned release status.

The Oktane release review separates endpoint discovery from the other announced features. For a discovery pilot, preserve that separation. Agent SSO being generally available is not evidence of endpoint discovery availability, and discovery finding an agent is not evidence that the agent’s next tool call is governed.

Read Microsoft’s sensor discovery as an identity pattern

Microsoft’s 19 October announcement expands Entra Agent ID with identity types for additional agent patterns, including local agents and previously unknown agents discovered through security sensors. It introduces Project Perception, sharing signals, context and a unified registry with Agent 365. The verified announcement does not establish prices or availability dates for these additions. Keep those unknowns visible rather than importing a date from another part of the product.

The wording matters because it connects previously unknown agents to an identity model. It does not establish that every sensor finding arrives with a complete permission history or an approved operational purpose. Ask what record is created, what context accompanies it and how the reviewer can connect it to an accountable person. These are proposed checks to make the discovery result usable, not additional Microsoft product claims.

The October Entra Agent ID review covers the wider announcement. For this narrower comparison, focus on the transition from observation to record. Request an example where an unfamiliar agent is found and reviewed, and keep that evidence separate from a demonstration involving an agent that was already registered.

Define the review after an agent is found

Build a review procedure before measuring a discovery demonstration. Decide who can accept a finding, who can assign an owner and who can require further investigation. Ask the supplier to show the information available at each step. An inventory entry should not become an implied approval simply because it has been given an identity or placed in a shared registry.

Use two deliberately different test cases. One should represent an agent the team expects to find, with a known owner and intended task. The other should represent an unfamiliar agent that needs review. Compare the resulting records and the administrative choices offered. This method tests the handoff between discovery and governance without assuming either vendor implements a particular quarantine or remediation workflow.

Also ask how repeated observations are handled and how a reviewer knows whether a record is current. Record the supplier’s answer rather than assuming that a discovery mechanism implies a particular reconciliation method. The important operational question is not just that an agent was seen, but whether the team can explain the finding later and connect it to a defensible decision.

Keep discovery separate from action governance

Discovery is one boundary in the control model. Registration, permission assignment and runtime authorization are others. The agent identity comparison helps place those boundaries in the broader provider discussion. For this evaluation, require separate evidence for each step. A discovery result should not be counted as proof of credential protection, human approval or the ability to stop work already in progress.

NewCore is an agent governance layer that runs alongside the identity provider a company already has. Its October award announcement states that it continuously discovers agents, including shadow agents no one registered. It also describes named accountable owners, task-scoped credentials, policy evaluation on every call and human approval for high-risk actions. That is relevant when the requirement extends beyond finding an agent to governing what it may do.

Compare those vendor-stated capabilities against the same bounded workflow rather than merging them with an identity provider’s release claims. Keep the final record organized around observation, identity, ownership and action decisions. The outcome should be a clear account of what each source establishes, what the demonstration proves and which gaps still require documentation or testing before a wider deployment.

Frequently asked questions

Does Okta describe how it finds coding agents?

Yes. Its NIST checklist describes a browser plugin and EDR integrations, and names Claude Code and Cursor as examples of coding agents it finds. That mechanism description should be read alongside the September release’s planned Q3 availability statement for endpoint discovery.

Are Microsoft’s sensor-discovered identities dated for release?

The verified October announcement describes the new identity patterns but does not establish an availability date or price for the additions. Do not infer those details from broader Entra Agent ID platform statements or from the announcement’s reference to Microsoft 365 E7.

What should follow a discovery finding?

Define a review that establishes what was observed, the responsible person and the intended next decision. Then request separate evidence for permissions and action controls. This is an evaluation recommendation, not a claim that every discovery product automatically provides those controls.