Aona FDSE service
Who checks security
in your AI rollout?
Aona reviews controls alongside your implementation team, with clear findings for launch.
Security already belongs
inside delivery.
Anthropic’s Frontier Deployed Engineer training takes engineers through a simulated enterprise deployment, including security review and handover. AWS describes its FDE model as embedded delivery that leaves customers able to operate independently. Both make production responsibility part of the engineering job.
Aona’s view: keep that responsibility with delivery, and add a clear route for challenging its conclusions. A new security title helps only when someone can inspect the evidence, record disagreement and escalate beyond the launch deadline’s owner.
The NIST AI RMF Playbook, GOVERN 2.1 and 2.3, suggests separate management for development and testing, transparent conflicts of interest and executive responsibility for AI risk decisions. It offers guidance, not a required organisation chart.
Assign decisions,
not just job titles.
Delivery team
Proposes architecture and access settings, implements approved changes, supplies configuration evidence and fixes failures. The customer’s data owner confirms which records each role may use.
Aona security
Agrees test criteria before execution, checks permitted and forbidden actions, and records results and limitations. We recommend proceed, restrict or pause, with unresolved findings visible to the accountable owner.
Customer owner
Approves the launch scope and any time-limited exceptions within their delegated authority. Assigns an operational owner for monitoring, incidents and access removal after the implementation team leaves.
Aona’s recommended operating model. In a smaller organisation, people may cover several roles. Make the reviewer’s reporting path and any conflict explicit.
Illustrative scenario
The employee moves.
Their AI access must follow.
A project assistant passed its launch checks. A team member then moves to another department and loses access to the project’s restricted document collection.
After the documented permission synchronisation, a controlled test confirms that the source blocks that account. Yet in a fresh conversation, the assistant returns a unique fictional fact from the test document. The previous launch result does not resolve this new finding.
Delivery investigates connector permissions, identity mapping and synchronisation with the data owner. The reviewer retests after the documented propagation process, with a fresh conversation and both a current member and the removed member. The customer decides whether to suspend the connection while the boundary is restored.
Restrict, then retest.
- Finding
- A removed project member can still retrieve the restricted test fact through the assistant.
- Immediate boundary
- Suspend the affected connection until revocation and permitted-access checks pass.
- Evidence to retain
- Account and group, source permissions, connector settings, query, result and test time.
- Customer decision
- Approve the narrower pilot, or pause if the collection is essential.
Make the decision
reproducible.
Aona recommends a short evidence packet that another person can follow: the tested configuration, representative roles, sample inputs, expected and observed results, unresolved findings, and the person responsible for each fix.
- 01Proceed. Agreed checks pass for the recorded scope. Name the monitoring owner and next review date.
- 02Restrict. Remove the affected source, action or audience and verify the smaller boundary. Record what remains unavailable.
- 03Pause. A material exposure remains, or essential evidence is missing. Record the condition that would allow a retest.
An accepted exception needs an accountable owner, reason, compensating control and expiry. “Could not test” remains an evidence gap, not a passed check.
Reopen review
when access changes.
A previous result applies to its tested configuration. Recheck the affected boundary when a connector is added, a new data class enters scope, a pilot expands to another role, or group membership changes.
Moving from reading documents to sending messages or modifying records needs new action-specific checks. Assign someone to assess relevant provider changes, incidents and failed controls, rather than waiting for the next annual review.
What “separate”
means at Aona.
Our AI Security Partner pilot provides a review workstream alongside your implementation team, focused on employees’ AI use and sensitive-data exposure. Aona also supplies security software, so this is not an independent audit opinion on Aona’s products. Agree the reviewer, evidence access and escalation path in the engagement. You approve configuration changes and launch.