The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines advisors that process and store AI-generated completion data.
Ingestion points: templates/AiAuditAdvisor.java and examples/good-ai-audit-advisor.java ingest ChatClientResponse and ChatResponse objects from external AI providers.
Boundary markers: The templates do not utilize specific delimiters for completion content; however, they focus on metadata extraction rather than content re-processing.
Capability inventory: The skill uses MeterRegistry for metrics, Slf4j for logging, and AiAuditLogRepository for database persistence.
Sanitization: The provided production templates (templates/AiAuditAdvisor.java) specifically sanitize the process by logging only UUIDs and token counts, explicitly avoiding the storage or logging of raw prompt and completion text to mitigate PII exposure.
[DATA_EXFILTRATION]: The configuration recommendations include enabling Spring Boot Actuator endpoints for metrics collection.
Evidence: SKILL.md suggests exposing health, prometheus, and metrics via management.endpoints.web.exposure.include. While standard for observability, these endpoints can expose internal service metadata if not properly secured with Spring Security in a production environment.
[DATA_EXPOSURE]: The skill includes an educational example of insecure practices.
Evidence: examples/bad-ai-audit-advisor.java demonstrates the risk of logging full prompt contents (request.prompt().getContents()) and using hardcoded pricing, which are explicitly marked as patterns to avoid for security and maintainability.

ai-observability