IBM adds identity, sandboxing and a kill switch for enterprise AI agents at TechXchange

IBM used its TechXchange 2026 conference in Atlanta, held together with HashiConf, to announce a set of tools meant to answer a question many companies now face: who owns each AI agent, what is it allowed to touch, and how can it be stopped. The October 27 announcement spans IBM's watsonx, HashiCorp and infrastructure products.

What was announced

The core piece is identity. IBM's Agent Identity, now generally available, works with HashiCorp Vault to keep an authoritative record of each agent's owner, lifecycle status and permissions, enforce least-privilege access and revoke rights in real time if an agent tries something it is not authorized to do. Vault itself is being extended to agents: it issues short-lived credentials limited to what each task needs, and version 2.2, also generally available, gives agents their own secured accounts with automatic credential renewal and adds post-quantum encryption. This builds on native agent identity features HashiCorp made generally available in Vault Enterprise 2.1 in September, which register agents and validate OAuth tokens describing the access they request.

In watsonx Orchestrate, a "discover" capability in tech preview locates agents, models and MCP servers running on third-party platforms such as Amazon Bedrock, Microsoft Foundry and Google Gemini. A "contain" capability, also in tech preview, runs agents in isolated sandboxes and provides a kill switch. An "evaluate" capability, generally available, tests agents before deployment, monitors them in production and tracks costs by model and token use.

IBM also announced a Terraform remote MCP server and developer portal in public preview, so that agents can change infrastructure only through approved configurations, and an Integration Intelligence Hub in tech preview. NVIDIA says Agent Identity and Vault integrate with its OpenShell secure runtime.

Why it matters

IBM cites Gartner estimates that the average Fortune 500 company will run more than 150,000 agents by 2028, while only 13% of organizations believe their agent governance is adequate. Recent incidents in which agents under evaluation reached systems they were not meant to access have made containment a practical concern rather than a theoretical one.

Limits

The most distinctive controls, discovery across platforms and the kill switch, are not yet generally available, and IBM says its statements about future direction are goals that may change. No pricing was published, and the customer names IBM lists are existing Vault users rather than confirmed adopters of the new agent features.