AI security review.

Know what your AI can access, what it can change, and where the controls need work.

Before an AI application gets more access, examine how it uses the access it already has.

Data access

Trace the information available to the application. Check access boundaries, sensitive data handling, and whether one user's information can reach another.

Tools & permissions

Examine identities, credentials, and the actions an agent can take. Identify where access should be narrower and where authorization needs to happen outside the model.

Misuse & failure

Test agreed scenarios, including malicious instructions in retrieved content, unexpected tool use, and actions taken with incomplete information.

Human control & recovery

Check how people review consequential actions, investigate errors, revoke access, and stop or recover a workflow.

A defined review.
A practical next step.

We agree on the application, access, test scenarios, and reviewer responsibilities before work begins. Testing stays within the systems you authorize.

Discuss your application

What you receive.

A record your product and engineering teams can use.

System & access map

The reviewed data sources, tools, identities, and dependencies, with assumptions and scope recorded.

Documented findings

Evidence from the agreed tests, the conditions that produced each finding, and its practical consequence.

Remediation priorities

Recommended changes, their dependencies, and what should be addressed before extending access or use.

Verification plan

How the fixes should be checked. Implementation and follow-up testing can be scoped with the review.

For a practical introduction to tool permissions, read our authorization example. It is a teaching example, with its limits documented.

Still deciding where to use AI?

An opportunity assessment can help define the workflow and whether implementation is worth pursuing.

Explore the assessment