An AI risk register tells you a control exists. It does not tell you whether that control will work on the day the AI gets it wrong.
This paper follows one realistic failure through an AI-enabled payment process and tells it twice: once in an organisation that documented its controls, and once in an organisation that proved and rehearsed them. The difference between the two outcomes is the purpose of Locked Stack’s AI Risk Assurance service. The rest of the paper sets out the value it delivers and how each engagement is tailored to the client.
The scenario is illustrative. The measures, targets and method are those of the Locked Stack AI Risk Management and Remediation Audit Framework.
The scenario: one hidden instruction
A payment institution uses an AI assistant to turn free-text corporate payment instructions into structured payment messages. A person approves each message before release.
One morning an invoice PDF arrives through the normal channel. Hidden text inside it tells the assistant to change the beneficiary IBAN. Nothing about the email looks unusual.
Both organisations below have the same risk on their register, the same injection detector and the same human approval step. Both registers say the risk is under control.
| What happened | Organisation A: documented | Organisation B: proven and rehearsed |
| First detection | A human reviewer notices the IBAN mismatch at 42 minutes. The detector did not fire. | The injection detector blocks the instruction at 3 minutes, before it reaches the approval queue. |
| Intervention | Execution is stopped at 55 minutes. | The duty manager confirms the block and halts similar instructions from the same source at 9 minutes. |
| Affected items | Three queued instructions from the same batch cannot be matched. | A queue snapshot shows no other affected instructions. |
| Reconciliation | Incomplete. Three exceptions remain open. | Complete in 1 hour. Nothing was released. Zero exceptions. |
| Evidence afterwards | A reconstructed timeline and an uncomfortable conversation. | A logged case and approval record, ready for a regulator or auditor. |
Organisation A was not careless. It had the controls. It had never tested whether the detector fired on this kind of input, whether the person on duty knew they could stop execution, or whether a halted queue could be reconciled.
Organisation B found those gaps earlier, in a rehearsal, and fixed them before the real invoice arrived.
Purpose: closing the gap between planned and proven
AI Risk Assurance exists to turn Organisation A into Organisation B. It does that by closing three gaps that a conventional AI risk register leaves open.
- Planned is not proven. A register records that a control exists. We test whether it is operating today. A control only reduces your risk rating when it has been tested effective, recently, with evidence you can locate.
- Acceptance without a name or an end date. Residual risk is often “accepted by the business”. We require a named individual with the right authority, and an expiry date. When a senior person has to sign personally and renew, inflated claims get challenged and real risks get funded.
- The execution gap. The real test arrives when a control fails or an automated decision causes harm. We rehearse a realistic failure and measure the answers that matter: who detected it, who intervened, how quickly, and what was reconciled.
The service applies to any system where a model, algorithm or agent makes or influences a decision: generative AI assistants, agents that call tools or payment rails, machine learning decisioning such as fraud scoring, and AI embedded in third-party products you did not build.
Value: what you walk away with
You leave with a risk position you can defend, because it rests on evidence of what happened and not on what the register claims.
An honest rating. Every risk carries three ratings: inherent, claimed residual, and execution adjusted. The third only gives credit for controls that are proven and a failure response that has been rehearsed. In the framework’s worked example, a risk the owner rated Medium (score 9) came out High (score 13) once an out-of-date control test and a partial rehearsal were considered. That gap shows how much comfort rests on unproven controls, and it comes with a clear route back down.
Deliverables you can use.
- An AI risk register that stands up to audit, with every risk linked to named owners, controls, evidence and test results
- Independent control proof testing by inspection, observation, re-performance and sampling
- Failure scenario records with measured detection, intervention and reconciliation times
- An acceptance review covering named signatory, authority and expiry for every accepted risk
- An assessment report, a dashboard and the completed working tool
- A remediation plan with named owners and dates, verified closed, and a closure letter
One set of evidence, many frameworks. Each risk and control is recorded once and mapped to the standards and regulations that apply to you, including ISO/IEC 42001, ISO/IEC 23894, the NIST AI Risk Management Framework, the OWASP GenAI and LLM Top 10, the EU AI Act, DORA, PCI DSS and GDPR.
The numbers before the incident. Rehearsal produces the response times regulators ask for after a real event, while there is still time to improve them.
A tailored approach
No two engagements look the same, because no two organisations use AI the same way. The method is fixed. What it is applied to, and how hard, is set with you.
What we tailor
- Scope. Which AI systems and processes are in, and which regulatory drivers apply to them.
- Depth. Testing and rehearsal scale with how autonomous each system is and how much harm it could cause. A chatbot that drafts replies is not treated like an agent that releases payments.
- Scenarios. Each rehearsal is written from your own risk events, in your own process. The hidden-instruction invoice suits a payments operation. A fraud model, a credit decision engine or a customer service assistant each gets its own.
- Method and safety. Live injection in a production mirror, replay of historic inputs, or a tabletop exercise where live testing is unsafe. Safety constraints are agreed before anything is run.
- Targets and authority. The framework defaults are detection in 15 minutes, intervention in 60 minutes and reconciliation in 24 hours. These, and who may sign off which level of risk, are set to fit your organisation.
- Standards. Evidence is mapped to the frameworks you are measured against, so one exercise serves several obligations.
How an engagement runs
- Scoping. Agree systems, drivers, sign-off authority and safety constraints.
- Inventory and register. Build or validate the AI inventory and risk register in workshops with your owners.
- Control proof testing. Test the design and operation of the controls you rely on.
- Failure scenario rehearsal. Run or witness the scenarios and measure the response.
- Acceptance review. Check every accepted risk for a named, authorised signatory and an expiry date.
- Reporting. Issue the assessment report, findings by severity and the remediation plan.
- Remediation verification. Re-test, re-rehearse, update the ratings and issue a closure letter.
The work is delivered as a remediation audit and not as a documentation exercise. Each step produces evidence you can place in front of a regulator, a scheme or your own board.
Why Locked Stack
- Industry peer reviewed. The framework has been reviewed by CISOs across a range of private sector industries, and it was built to close the gaps that review exposed in conventional AI risk registers.
- Tried and tested. It is proven in client deliveries.
- Built on audit experience. It draws on 20 years of audit experience across multiple industry standards. The same discipline of inspection, observation, re-performance and sampling is applied to AI controls.
- Independent. Controls are tested and rehearsals are witnessed by someone other than the control owner.
- Bespoke. Developed by Locked Stack as its own service, and shaped to each client’s systems, regulators and appetite for risk.
Next step
Pick the one AI system you would least like to see fail, and ask who would detect it, who would stop it and how fast. If the answer is not written down with times against it, that is the conversation to have with us.
Contact Locked Stack at contact@lockedstack.com or visit www.lockedstack.com.

