The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides architectural templates that ingest untrusted data from LLM inputs and outputs into a logging pipeline.
Ingestion points: The withAuditLogging middleware in SKILL.md reads params.messages and response.content directly from the AI transaction.
Boundary markers: The code does not implement specific delimiters or warnings to treat ingested content as untrusted text within the log objects.
Capability inventory: The skill includes implementation logic for persistent storage via @prisma/client and custom storage classes.
Sanitization: While the logger uses sha256 hashing to protect privacy, the inputSummary, outputSummary, and metadata fields are stored as plain text, creating a surface for injection if logs are viewed in unsanitized dashboards.

ai-audit-logging