IBM adds identity and access controls for enterprise AI agents

The rise of non-human identities

As businesses begin to deploy AI agents that can perform work across different systems, these agents are becoming a new class of non-human identity. Unlike a human user who logs in with a password, an agent needs a way to prove who it is and what it is allowed to do. IBM is addressing this by creating a framework for agent discovery and verified identities.

r>

Currently, many organizations struggle with 'agent sprawl,' where they lose track of which agents exist and who owns them. This lack of visibility makes it difficult to manage permissions or audit actions across different platforms and clouds. IBM's new capabilities aim to provide a single place to discover and secure these agents regardless of where they were originally built.

By establishing an authoritative record for each agent, companies can see its lifecycle status and the specific authority it has been granted. This move is designed to help organizations delegate work with defined accountability. It ensures that every action taken by an AI can be traced back to a specific owner and a set of approved permissions.

Securing credentials with Vault

A major security risk for AI agents is the use of long-lived, overprivileged credentials. If an agent has a permanent key to a database and that key is compromised, it could lead to significant data breaches. To solve this, IBM is extending HashiCorp Vault to issue secured, short-lived credentials that expire automatically.

These credentials are designed to follow the principle of least-privilege, meaning an agent only gets the specific access it needs for a specific task. If an agent attempts to perform an action outside of its approved permissions, the system can deny the request and revoke the session in real time. This creates a dynamic security boundary that adjusts based on the agent's current activity.

The updated Vault 2.2 also includes features like automated credential renewal and post-quantum encryption to protect against future threats. By managing these secrets centrally, IT teams can reduce the time that exposed credentials remain usable. This approach allows companies to treat AI agent security with the same discipline as their most critical internal systems.

Sandboxing and kill switches

One of the primary concerns for enterprise leaders is a 'rogue' agent causing unintended damage or spreading mistakes across a business. To prevent this, IBM's watsonx Orchestrate now includes a 'contain' capability. This feature runs agents in isolated sandboxes, ensuring that one agent cannot interfere with others or disrupt the wider infrastructure.

In addition to isolation, the platform provides teams with a manual kill switch to shut down any agent immediately. This is a critical safety feature for organizations that are still learning how to interact with autonomous systems. It gives human operators ultimate control over the AI's ability to execute tasks in real-time.

The platform also includes an 'evaluate' feature to test the quality and accuracy of agents before they are deployed. Teams can monitor for 'drift,' which happens when an agent's performance changes over time or it begins to act outside of its intended parameters. This allows for better cost tracking and performance monitoring before an agent goes live.

Connecting to existing systems

For AI agents to be useful, they need to understand the existing business context, such as how systems connect and what APIs are available. IBM is using its Integration Intelligence Hub to provide this knowledge through a Unified Context Engine. This helps agents solve problems like stuck messages or deploy APIs using the company's existing infrastructure.

Instead of forcing companies to rebuild their entire tech stack for AI, this approach allows agents to work on the systems they already depend on. The system uses skills, approval policies, and audit controls to keep these actions in check. It bridges the gap between the AI's capabilities and the company's actual data and applications.

The solution also integrates with NVIDIA's secure runtime software to provide a protected environment for execution. This combination of verified identities, short-lived credentials, and policy-governed runtimes helps enterprises scale AI agents safely. It ensures that every action taken by an agent is traceable, limited, and reviewed by the organization.