Read the availability section, not the headline
The announcement bundles four ideas: discover every agent, give each one a verified identity, limit its access, and evaluate it. Read as one launch, it sounds like a single product a team can buy this week. The availability section at the bottom of IBM's own release separates it into three states.
Shipping now are the identity and credential parts. Agent Identity is generally available, and so is Vault 2.2. The cross-platform discovery and containment capabilities are specifically in technical preview. Discover, which connects to Amazon Bedrock, Microsoft Foundry and Google Gemini to find agents, models and MCP servers, is a tech preview. Contain, the isolated sandbox with a kill switch, is a tech preview as well.
That split decides who this launch is for today. A team that already knows its agents can use the available parts. A team whose problem is unknown agents on three rival clouds is being shown a preview of its answer, with no date in the release for when it becomes a product.
What the identity record actually does
The concrete mechanism is a record, not a model feature. Agent Identity, used with Vault, establishes who owns an agent, whether it is active, and what authority it has been granted. That record feeds Vault's credential controls, built to enforce least-privilege access: an agent gets the access its job needs, as credentials that expire and can be revoked.
The revocation claim is specific, and worth keeping in its own terms. IBM says Agent Identity, used with Vault, can enforce granted permissions and revoke access in real time when an agent attempts an unauthorized action. That is a claim about identity and credential enforcement, not proof that the system can detect every harmful action an agent might take while operating within its approved permissions. The release does not provide an independent enforcement test or specify every runtime condition under which revocation occurs.
Two gaps sit in the same paragraphs. The credentials are called short-lived throughout, and no lifetime is ever given. Vault 2.2's post-quantum encryption is named without detail about which operations it covers. Both are buyer questions, not specifications.
The 150,000 figure is a forecast doing sales work
IBM opens its case with Gartner: by 2028, the average Fortune 500 enterprise will use over 150,000 AI agents, while only 13% of organizations believe they have the right governance in place. The first number is a prediction about 2028, not a count of anything running today. IBM's release does not provide an operational definition of an agent for this forecast, and a definition is what would make the number testable.
The 13% figure records what organizations believe about their own governance, a survey answer, not an audit result.
The underlying problem is still real. Recent authorized security testing covered by this site illustrates why access boundaries matter: Anthropic confirmed limited access to internal sales, HR and IT systems through third-party software, while disputing aspects of the researchers' account. The confirmed access did not include Claude conversations, source code or model weights. An agent holding an overprivileged credential presents a separate, continuing risk. The forecast does a different job: it sizes a market in 2028 so that a control product bought in 2026 looks proportionate. The risk is present tense. The 150,000 is not.
Where this fits, and what to ask first
IBM is not the first vendor to sell agent identity. RSA launched an agent identity product on October 7, Microsoft added agent risk controls to its E7 tier on October 19, and 1Password began letting agents sign in without passwords on October 22. IBM's distinction is where the identity record sits. For organizations already using HashiCorp Vault to manage secrets and credentials, Agent Identity can connect agent ownership and permissions to an existing access-control system. That may reduce the need for a separate identity layer, although the release provides no adoption figures or deployment-cost comparison.
The useful evaluation questions follow the release's own silences. How long is short-lived, in a number? What does Agent Identity cost, and per what? When does Discover leave tech preview? IBM's release answers none of these. Until it does, the honest summary is a shipping identity layer, a preview discovery layer, and a forecast where a price should be.
Sources
- [1] IBM, "IBM Introduces Identity and Control for AI Agents, Securing Frontier Intelligence for the Enterprise", October 27, 2026, PR Newswire (primary release; all product definitions, the Dinesh Nirmal and John Fanelli quotes, the Gartner figures and the full Availability split)Read source
- [2] IBM release as syndicated by Morningstar, October 27, 2026 (full text read first-hand, including the Availability section and footnotes)Read source
- [3] Unite.AI, "IBM Brings Governed Identity and Credential Controls to AI Agents", October 27, 2026 (secondary account; adds that the cited Gartner press release is dated April 28, 2026)Read source
- [4] RuntimeWire, "IBM gives AI agents identities before they reach enterprise systems", October 26, 2026 (secondary account; notes IBM Newsroom as its primary source and the cross-platform positioning)Read source