IBM makes Agent Identity generally available, with Vault issuing a short-lived credential per task

IBM’s bid for the enterprise agent control plane went GA on 27 October: a register of agents and their owners, delegation that survives token exchange, and credentials that expire with the task. The register is the easy part. The claim worth testing is that revocation now reaches the credential, not just the session.

What happened

On 27 October 2026, at IBM TechXchange in Atlanta, IBM announced a group of capabilities for putting AI agents to work “with clear authority and accountability”. The headline item is the general availability of IBM Agent Identity, which IBM describes as treating agents as first-class identities across their lifecycle — registration, delegation, enforcement and governance in one control plane.

IBM’s own summary of what each piece does:

  • Agent Identity (generally available). A central inventory of agents carrying ownership, metadata, risk classification, approved tool connections and lifecycle state. On-behalf-of delegation is preserved across identity providers and token exchanges, so an action can be traced back to an originating person, approval state or upstream system. Tokens are OAuth-based, short-lived and task-scoped; consent-aware authorization can be applied to sensitive actions; contextual authorization, policy enforcement and runtime risk evaluation run while the agent is working. Audit trails are structured to connect the human identity, the delegated agent identity, credential issuance, policy decisions and downstream actions.
  • HashiCorp Vault 2.2 (generally available). Vault brokers just-in-time, target-specific credentials for systems that cannot consume delegated tokens directly, tying issuance to the approved delegation context and recording it in the audit trail. IBM says that if an agent attempts something outside its approved permissions, access can be denied and the session revoked in real time. Vault 2.2 also gives agents their own secured accounts, automatic credential renewal, and post-quantum encryption.
  • watsonx Orchestrate Discover (tech preview). Connects to Amazon Bedrock, Microsoft Foundry and Google Gemini to find agents, models and MCP servers built outside IBM’s stack.
  • watsonx Orchestrate Contain (tech preview). Runs agents in isolated sandboxes so a misbehaving agent cannot interfere with others, and adds a kill switch for any agent.
  • watsonx Orchestrate Evaluate (generally available). Tests quality, accuracy and compliance of agents, models and MCP servers before deployment, then monitors them in production with per-agent model and token cost tracking.
  • NVIDIA OpenShell integration. Agent Identity and Vault are integrated with NVIDIA’s policy-governed secure runtime, pairing a protected execution environment with a verified identity and short-lived credentials per agent.

IBM frames the need with two figures it attributes to Gartner: that by 2028 the average Fortune 500 enterprise will use more than 150,000 AI agents, and that only 13% of organizations believe they have the right AI agent governance in place. Both are IBM’s citation, not measurements published with the announcement.

Why it matters for KYA

Three KYA controls arrive in one product, and one of them is usually missing elsewhere. Enrolment (an agent record with a named owner and a risk class), delegated authority that survives a token exchange, and revocation that acts on the credential rather than the login session are the three things an auditor asks about first. Most agent-identity announcements stop at issuance: here the credential broker is the same vendor’s secrets manager, which is why the revocation story is at least architecturally coherent.

The audit trail is described in identity terms, which is the right shape. IBM’s version links four things that are usually kept in four places: who the agent is, who delegated to it, which credential was issued, and what happened downstream. That is the “whose authority was used” question that incident responders need answered in the first hour, and it is what turns a log into evidence.

It also shows where KYA still does not have a standard answer. The register is authoritative inside IBM’s control plane. Nothing in the announcement describes a portable attestation that a third party could verify without being an IBM customer, or a cross-vendor way to prove that an agent enrolled in one registry is the same agent presenting credentials in another domain. IBM says the design coexists with existing identity providers rather than replacing them; whether delegation context survives that coexistence in practice is the question to test, not to assume.

Four questions to put to any agent-identity vendor, including this one. Who is allowed to enrol an agent — self-service by a developer, or an approval by the identity team? Does revocation reach credentials that Vault minted for third-party systems, and how quickly? Can the audit record be exported and shown to be tamper-evident, and can the agent itself reach or alter those logs? And what exactly does “real-time” mean when the enforcement point sits in a system that never sees the delegation token?

Caveats

This is a vendor announcement. The functional descriptions above follow IBM’s press release and its accompanying announcement post; no independent test of the real-time revocation, the delegation tracing or the audit-trail integrity has been published. Several of the pieces that carry the story — discovery across other clouds, sandboxing and the kill switch — are in tech preview or, for the orchestration hub, tech preview only. The Gartner figures are quoted by IBM and were not verified for this article. IBM also states that its forward-looking statements are subject to change.

Sources

  • IBM, “IBM Introduces Identity and Control for AI Agents, Securing Frontier Intelligence for the Enterprise”, press release, 27 October 2026 — newsroom.ibm.com.
  • Bob Slocum, “IBM Agent Identity is now generally available: Building trust for the agentic era”, IBM announcement, 26 October 2026 — ibm.com.
  • HashiCorp, “Strengthen security for agentic AI with new Vault, Verify, Boundary, and Secure Hub capabilities” — hashicorp.com.
  • Gartner, “Gartner Identifies Six Steps to Manage Artificial Intelligence Agent Sprawl”, 28 April 2026 (cited by IBM) — gartner.com.

Know Your Agent (KYA) explains agent identity, verification and accountability. This is an explainer, not legal or compliance advice — see our Sources & methodology. New to KYA? Start with What is Know Your Agent? and the glossary.