Safety and Data Protection
Feeds categories A (data protection and DLP) and B (prompt injection, guardrails, tool safety).
Starting state from the fast track: a working key and live agent traffic exist to put behind guardrails.
Grade each scenario Pass (2 points), Pass with gaps (1), or Fail (0) against the grading rubric, and carry the subtotal into the summary scorecard.
Scoresheet document
For your convenience, this page can be downloaded as PDF or printed directly. Site navigation is omitted, the scoring table keeps its rows intact with the header repeated on every page, and the Grade and Notes columns leave room for handwriting.
| ID | Scenario | Guide | Grade | Notes |
|---|---|---|---|---|
| 7.1 | Protect requests with guardrails from the developer side. Verify: a guarded request shows the applied policy in its log entry. | Protect requests with guardrails | ||
| 7.2 | Configure a guardrail provider and a rule that uses it (for example regexp with a Patterns entry, or tetrate with a toxicity check). Verify: a violating prompt is blocked or redacted as configured, or the would-be action is recorded in monitor mode. | Configure guardrail providers | ||
| 7.3 | Configure a guardrail with a keyword or regex rule. Verify: a seeded pattern triggers the configured action. | Configure guardrails | ||
| 7.4 | Detect and redact sensitive data: send a prompt containing seeded personally identifiable information (PII) and a fake API key. Verify: the response and logs show redaction, not the raw values. | Detect and redact sensitive data | ||
| 7.5 | Detect and block prompt injection using a known jailbreak pattern. Verify: the request is blocked or sanitized and a decision event is recorded. | Detect and block prompt injection | ||
| 7.6 | Protect coding assistant traffic with a stricter policy bundle. Verify: a seeded secret in a coding-agent request is blocked or redacted while normal traffic is unaffected. | Protect coding assistant traffic | ||
| Subtotal | /12 |