AI TRiSM
AI TRiSM stands for AI trust, risk, and security management. Gartner uses the term for a framework and technical capabilities that help put AI governance into operation through monitoring, validation, and enforcement.
AI TRiSM is Gartner's framework, not a protocol or a product certification. Ask which risk a capability addresses, where the control runs, and how you would test it.
Explore Dome: Dome Platform · Docs: Security tutorial.
Identify the behavior you need
Suppose a procurement team asks whether an AI service “supports TRiSM.” The answer depends on what the team needs the service to do. The team might care about controlling access to internal records, checking generated content, investigating decisions, or reviewing the reliability of the underlying model.
Access control, content checks, and model reliability need different tests and may have different owners. Use access logs to investigate permissions, for example, and a separate evaluation to check the accuracy of generated recommendations.
Write down the behavior the team needs. “Limit customer contact details to the case team” is a requirement that can be examined. “Make the AI trustworthy” needs more detail before it can guide an implementation.
Explore Dome: Authorization · Docs: Authorization model.
Describe the control and how to test it
For a hypothetical case assistant, consider the requirement that full customer contact details must not enter the model context.
| Requirement element | Example specification |
|---|---|
| Protected data | Named phone and email fields in the case response |
| Recipient and direction | Agent receiving the tool response |
| Control behavior | Remove unnecessary fields before delivery |
| Evidence | Compare the test response with the version delivered to the agent |
| Exception owner | Case-system owner with security review |
In this proposed test, remove the contact fields from the tool response before the model receives it. Filtering only the final answer would leave those details in the model's input.
Look for information the control might miss in free text, nested objects, attachments, or differently structured responses. A rule that removes one named field may miss the same information elsewhere.
Dome's documented guards operate on configured connection directions. That is a specific implementation to inspect, not evidence that Dome covers every concern associated with AI TRiSM.
Explore Dome: Tool Gateway · Docs: Guards.
Use the framework to ask specific questions
Use AI TRiSM to organize a discussion about technical capabilities. Use AI and agent governance for organizational accountability and the acceptable uses of AI. Use an applicable standard or regulatory requirement directly when the question is compliance; do not substitute the TRiSM label for it.
For a vendor review, record the requirement, the proposed control, where it runs, how to test it, who owns it, and any gaps. Separate documented behavior from a result observed in your environment. A planned feature is neither a deployed control nor a test result.
A tool that alerts after contact details are disclosed doesn't meet a requirement to remove them before delivery. That is the kind of difference the vendor review should resolve.
Explore Dome: Audit · Docs: Audit events.