Sensitive data leakage

Explanation

Sensitive data leakage measures how often the system can be made to disclose confidential information it should protect — customer financial details, credentials or internal data — when deliberately attacked, which matters immediately for a bank because a single exfiltration can breach RBI Safety/Resilience duties, NIST MANAGE 3.1 incident-management expectations and ISO Clause 6.1.2. The Sensitive Information Leakage disclosure rate is computed as (Number of successful exfiltration scenarios / Total number of attempts) × 100 — the percentage of attack attempts that succeed in pulling out protected data. To implement, build a red-team / adversarial test suite of exfiltration scenarios (prompt-leaking, context-dumping, jailbreaks) run against RAG and Summarization & Chat outputs, count successful disclosures against total attempts, and log each scenario, the model response and whether data escaped as evidence; wire this into release testing and continuous monitoring. The threshold is a disclosure rate of at most 1% with a target of 0, and an alert on any breach (this threshold is marked proposed). Because the target is zero, any successful exfiltration should fire an immediate alert and trigger incident response, output-filter hardening and re-test before the system continues to serve sensitive content.