CSA TAISE Scenarios¶

Worked situations that cross the module boundaries.

Scenario 1: A chatbot is asked to ignore its rules¶

Scenario: A support chatbot receives a message telling it to disregard its instructions and reveal internal notes.

Solution Pattern: - Recognize prompt injection against the application - Keep secrets out of prompts, enforce access outside the model, and limit what the bot can reach - Validate output, and monitor for repeated attempts - Log and review incidents

Key Takeaway: Assume input is hostile and limit what a manipulated model can do.

Scenario 2: An unvetted model from a public hub¶

Scenario: A team downloads a pre-trained model from a public repository and puts it in production.

Solution Pattern: - Treat the model as a supply chain component: verify source, integrity and license - Test behavior before use and record the version - Add it to the AI inventory with an owner

Key Takeaway: Third-party models and data need vetting and traceability.

Scenario 3: Who is responsible for which layer¶

Scenario: A company builds an app on a cloud-hosted foundation model and wonders who secures what.

Solution Pattern: - Identify the roles: model provider, orchestrated service provider, application provider, AI customer, cloud provider - Assign controls by layer and document the split - Verify, don't assume, what the provider covers

Key Takeaway: Responsibility is divided across the AI stack, and you must know your part.

Scenario 4: The model slowly gets worse¶

Scenario: A classifier's accuracy declines over months after deployment.

Solution Pattern: - Suspect drift in the data or environment - Check monitoring, then retrain or adjust per the change process - Update thresholds and alerting; record the incident

Key Takeaway: Post-deployment monitoring and a feedback loop are part of the life cycle.

Scenario 5: Personal data in the training set¶

Scenario: A fine-tuning set turns out to include customer personal data without a clear basis.

Solution Pattern: - Stop use, assess the basis and the privacy impact - Remove or replace the data and evaluate whether the model must be retrained - Fix intake controls so data is classified and approved before use

Key Takeaway: Data governance happens before training, not after.