1. Map supplier dependencies and contractual responsibility
For prompt-injection boundary: retrieved instructions, approved tools, output context and user consent, list each external supplier that can store data, issue credentials, process transactions or support recovery. Ask which organisation owns the policy and which can actually enforce it. A contract that names a control without specifying the executing party is insufficient.
Verify an exception for the “1. Map supplier dependencies and contractual responsibility” chapter of Supplier risk review guide for agentic AI and prompt-injection defences: ask a second reviewer to try the forbidden route and explain the rejection evidence while reviewing an assistant acting on an approved source document. Assign the AI product owner to test prompt-injection attempts and unauthorised write-back, retain a prompt, source excerpt, reviewer decision and corrected output and agree when to disable the connector and revert to a human-reviewed draft. Record the actual result and any unresolved condition;
The service owner should ask AI product owner and information governance lead to walk one fictional request involving an assistant-generated draft using a restricted knowledge article from intake to sign-off. Capture the starting permission, the system that contains the authoritative record and the moment the action becomes consequential. Introduce “the response cites a file the operator is not authorised to open” as the adverse condition, but do not tell the operator which control is being tested. The observer records each decision, any manual override, the point at which work is paused and whether Technology & SaaS can continue safely while the exception is resolved. Use fictional identifiers so the rehearsal does not create a real privacy or service incident. For the map supplier dependencies and contractual responsibility step, retain a before-and-after sample and ask a second reviewer to compare the actual record with the expected approval state. the responsible team can be able to explain any mismatch without access to the developer who configured the workflow. Where the evidence is ambiguous, mark the case unresolved and request a bounded retest; a completed checklist is not the same as a verified outcome. Document the result for “Supplier risk review guide for agentic AI and prompt-injection defences” under the 1. map supplier dependencies and contractual responsibility control, with the responsible owner and a follow-up check after the correction is implemented.
