IBM's answer to the rogue-agent problem is to give every AI agent its own identity, short-lived credentials and a kill switch. Announced October 27, 2026 at TechXchange, the package extends HashiCorp Vault and watsonx Orchestrate to agents, and parts of it are already generally available.
What IBM announced
On October 27, 2026, IBM (NYSE: IBM) announced new capabilities at TechXchange for controlling AI agents that act across business systems. The centerpiece is an extension of HashiCorp Vault, its secrets-management product, to AI agents. Vault now issues secured, short-lived credentials that limit each agent to what it needs, and it renews those credentials automatically. Vault 2.2 also gives agents their own secured accounts and adds post-quantum encryption.
Alongside Vault, IBM introduced Agent Identity. According to the release, it creates the authoritative record for each agent: who owns it, its lifecycle status and the authority it has been granted. If an agent tries to act outside its permissions, access can be denied and the session revoked in real time.
The watsonx Orchestrate layer
Orchestrate covers discovery, containment and evaluation. Discovery is meant to find agents built on Amazon Bedrock, Microsoft Foundry and Google Gemini, and it extends to models and MCP servers. Containment runs agents in isolated sandboxes and includes a kill switch to shut down any agent. Evaluation tests the quality, accuracy and compliance of agents, models and MCP servers before deployment, then monitors them in production, with cost tracking for each agent's model and token use. IBM says Vault and Agent Identity work with Orchestrate to authorize, audit and revoke agent actions. NVIDIA is integrating both with its OpenShell secure runtime.
What is live and what is still preview
Read the availability section before getting excited. Agent Identity, Vault 2.2, the Orchestrate evaluate capabilities and IBM Concert are generally available. Secure Hub and the Terraform developer portal and remote MCP server are in public preview. The Orchestrate discover and contain features, plus IBM Integration Intelligence Hub, are in technology preview. The release discloses no pricing. The sandboxes and the kill switch, the two features a solo operator would actually reach for, are still in preview.
The governance gap, by IBM's own numbers
IBM cites Gartner for two figures. By 2028, the average Fortune 500 enterprise will use more than 150,000 AI agents. Only 13 percent of organizations believe they have the right AI agent governance in place. Adoption is going to outrun control by a wide margin, and the release is built for the companies that already know it. Vault's non-human identity customers named in the release include ABN AMRO, Adobe, BlackRock, Canva, Starbucks and Target, which tells you who this is for.
What it means for solo operators and marketers
If you run a one-person shop, you will not buy Vault this quarter, and you should not. The design pattern underneath it is still the one to demand from every agent tool you touch. Each agent should have its own identity instead of borrowing yours. Its permissions should be narrow and time-limited. You should be able to see what it did and shut it off fast. Ask your agent vendor three questions: whose credentials does it use, how long do they last, and where is the kill switch? If the answers are "yours," "forever" and "in a support ticket," you have a liability, not a workflow.
Marketers have a specific exposure. Ad accounts, email platforms and analytics properties are exactly the systems an over-permissioned agent can damage in an afternoon. After managing more than 3,500 Google Ads accounts since 2006, I can tell you shared logins are where accounts go to die. An agent with your password is the same problem with less human judgment in the loop.
Our take
IBM is solving the right problem. An agent running on your login is a liability wearing a productivity hat, and a per-agent identity with short-lived credentials and an instant revoke is how you fix it. My skepticism is about timing. Enterprise vendors love announcing agent governance at a conference, and the features that matter most are still in preview. The most honest number in the release is Gartner's 13 percent, and it should worry you more than the 150,000-agent forecast. Having co-founded Urchin, which became Google Analytics, I have always believed the tools people trust are the ones that show what happened. Agent governance needs the same audit trail, and the vendor that makes a solo operator's agent log readable in plain English will do more for this market than any Vault-scale feature. We are now writing badge policies for software interns who never sleep, which is either progress or a very expensive joke.
Sources
IBM Newsroom, "IBM Introduces Identity and Control for AI Agents, Securing Frontier Intelligence for the Enterprise" (October 27, 2026): newsroom.ibm.com
PR Newswire via Manila Times, same release (October 27, 2026): manilatimes.net