Make the permissions, data flows and approval boundaries of an AI system visible before it starts taking action.
Discuss Your ProjectThe decision to resolve
Follow the information and actions through the entire system, then check that its controls hold when the input or workflow changes.
The relevant boundary includes retrieved documents, connected tools, credentials and logs. We identify which person or agent is acting, what it can reach and where information goes. Read access and permission to change a record are reviewed separately.
Important actions need a defined approver and enough information for a decision. Software checks the proposed action before it reaches that person. A model’s instructions or confident answer cannot grant the authority to make a change.
Testing follows the scoped workflow: untrusted instructions, requests beyond the agent’s role and attempts to misuse connected tools. Findings need supporting evidence. We distinguish an observed weakness, a recommended fix and a control that has been retested.
Governance becomes useful when your team can inspect the permissions, decisions and evidence around a specific workflow.
Define the identity behind each connection, its owner and permitted resources. Agree how access is changed or revoked, including the separation between reading, preparing and applying a change.
Identity and access boundaries
Review the sources, model requests, outputs and retained records. Define what information is necessary and which boundaries must remain in place as data moves between components.
Data flow and handling requirements
Specify which actions wait for a named person, what that person sees and how the decision is recorded. Retain observable proposals, tool requests and results for investigation.
Approval paths and reviewable records
Agree the testing surface and evidence standard. Record unresolved limitations and decide which changes require another review before the system’s access or responsibilities expand.
Findings, fixes and retest evidence
Threat modelling, permission boundaries, testing and operational evidence.
Explicit access to business systems, with controlled credentials and actions.
Understand the runtime and controls surrounding a model or agent.
We work with the system owner and your security team to agree access, a suitable test environment and the assessment boundaries. The review follows the tools and data the system actually uses.
The workflow, connected systems, trust boundaries and actions included in the assessment, with explicit exclusions.
Reproducible evidence, the affected control and recommended changes, with retest results where included.
Identity owners, approval responsibilities and the changes or incidents that should trigger another review.
Yes. A review can start with the existing architecture, workflow and available test environment. The access, testing boundaries and expected evidence are agreed before the assessment.
No. We can implement controls, document findings and map evidence to agreed requirements within scope. Legal interpretation, formal assurance and certification remain separate. Scoped testing cannot establish the absence of every risk.