IBM adds identity, credentials and a kill switch for AI agents: what is ready now and what is still preview
At TechXchange on 27 October IBM announced tools to give AI agents their own identities, short-lived credentials, sandboxes and a kill switch. Some parts are generally available; others are still in preview.
- Author
Umair Mateen- Published
- Reading time
- 4 min
- Category
- AI
On 27 October 2026, at its TechXchange conference in Atlanta, IBM announced a set of tools for finding, identifying, controlling and tracing AI agents across a company's systems. The pitch is simple: treat each agent like an employee or service account, with its own identity, limited access and a record of what it did.
An "AI agent" here means software that uses a language model to take actions, such as calling APIs, changing records or deploying code, rather than only answering questions.
What IBM announced
According to IBM's press release, the main pieces are:
- HashiCorp Vault 2.2 (generally available). Vault already stores and rotates secrets such as API keys and passwords. Version 2.2 gives agents their own secured accounts, issues them short-lived credentials limited to what each agent needs, renews those credentials automatically and adds post-quantum encryption.
- Agent Identity (generally available). Creates a record for each agent covering its owner, lifecycle status and what it is allowed to do. It shares that context with Vault so access can be revoked in real time if an agent tries something it should not.
- watsonx Orchestrate gains three capabilities. Evaluate (generally available) tests agents, models and MCP servers for quality, accuracy and compliance before launch, monitors them in production and tracks cost by model and token use. Discover (tech preview) finds agents built on other platforms, including Amazon Bedrock, Microsoft Foundry and Google Gemini. Contain (tech preview) runs agents in isolated sandboxes and adds a kill switch to shut any agent down.
- Secure Hub (public preview), a central view of health, usage and security across Vault environments.
- Terraform updates for infrastructure teams: policy controls (generally available), plus a developer portal and a remote MCP server (both public preview) so people and agents can make infrastructure changes through governed workflows.
- IBM Integration Intelligence Hub (tech preview), which gives agents context about existing integrations so they can fix issues such as stuck messages, under skills, approval policies and audit controls. IBM also listed IBM Concert (generally available) and IBM Z Atlas (public preview) for operations and mainframe teams.
MCP, the Model Context Protocol, is an open standard for connecting AI models to tools and data. Its appearance across these products reflects how widely it is now used to plug agents into business systems.
The release also notes that Agent Identity and Vault integrate with NVIDIA's OpenShell secure runtime, building on IBM's September support for NVIDIA's Open Agent Safety Platform, when Agent Identity was still in public preview.
Who should care
This matters most to mid-size and large organisations that already have agents running, often built by different teams on different platforms. IBM cites Gartner research saying only 13% of organisations believe they have the right agent governance in place. Whatever the exact figure in your company, the common gaps are the same: nobody has a full list of agents, agents share long-lived API keys, and there is no fast way to stop one that misbehaves.
Teams already using HashiCorp Vault or Terraform will find the generally available parts the easiest to adopt, since they extend tools those teams know.
Limits and caveats
- The cross-platform discovery and the kill switch, arguably the most useful pieces for companies with mixed agent estates, are tech previews, not production features.
- IBM's release gives no pricing and no dates for when preview features become generally available.
- These are IBM's own descriptions. Independent testing of how well discovery finds agents on other clouds is not yet available.
What to do this week
- Make an agent inventory. List every agent or automation that uses an AI model, who owns it and what systems it can touch. You need this whichever vendor you use.
- Replace shared, long-lived keys. Give each agent its own credentials with the narrowest access possible, and set them to expire.
- Define a stop button. Agree how you would disable any agent within minutes, and test it once.
- If you are an IBM or HashiCorp customer, ask your account team which of these features your current licences include and when the previews are expected to reach general availability.
The DevDrift helps teams put AI automation into their workflows with approvals, scoped access and audit trails.
