# AAISM interview scorecard

Candidate: ______  Interviewer: ______  Date: ______

| # | Question | Listen for | Score 1 to 4 |
|---|---|---|---|
| 1 | You learn the sales team has been pasting customer records into a public chatbot. What do you do in your first week? | Work with legal and privacy on where the data went, then find out why the rule failed: no approved tool, no training, no owner, nothing enforcing it. Fix that layer and bring the use into the inventory. | |
| 2 | How would you build an inventory of the AI we run? | Counts use cases as well as models; looks past data science to AI features switched on in licensed software, tools paid on corporate cards and notebooks on laptops; names an owner for each; classifies by sensitivity and risk. | |
| 3 | A model behind a critical process turns out to be poisoned. How does the continuity plan cope? | Integrity next to availability; a return to a version proven known-good and validated before it carries live traffic; a manual fallback people can still run; AI dependencies and providers listed in the business impact analysis. | |
| 4 | The data science lead says the model is 97% accurate, so the risk is low. How do you respond? | Accuracy is a test-set measure, not a likelihood. Asks about volume, which way the errors fall, results by segment and the controls that catch errors before they land. | |
| 5 | Who should accept residual AI risk, and on what terms? | The business owner with the authority under the governance charter, within the board's appetite, in writing, for a set period. Security assesses, advises and designs controls, and does not accept risk on the owner's behalf. | |
| 6 | A vendor switches on an AI assistant inside a platform we already license. What happens next? | Reopen the assessment for the change; find where the data goes, fourth parties included; check data-use and training terms; ask for evidence that covers the new feature, and contract terms that require notice before new AI processing. | |
| 7 | How would you stop an agent that can send email from being turned against us? | Least agency: narrow tools, a scoped identity, human approval for external messages and other irreversible actions, caps on steps, spend and rate enforced in code outside the model, and a kill switch. | |
| 8 | Can a better system prompt fix prompt injection? | No. The model reads instructions and data as one stream of text. Assume some injection succeeds, limit what the model can do, enforce access outside the model and validate output before anything acts on it. | |
| 9 | How do you know a guardrail is working? | Evidence from operation: seeded attacks caught, attack success and false positive rates tracked against agreed thresholds, alerts acted on, and a check that the monitor still sees traffic. | |

Source: https://credentialpress.com/guides/aaism-interview-questions
