The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill guides the agent to process and store application events that may include untrusted data from users or external systems.
Ingestion points: Data enters the system via auditEnricher (extracting request headers, IP addresses, and user agents) and call-site instrumentation using log.audit and auditDiff, which capture actor information and resource state changes (e.g., in SKILL.md).
Boundary markers: The skill does not define specific delimiters to isolate untrusted content from the audit system's processing logic, although it does advocate for redaction.
Capability inventory: The skill possesses capabilities to write this data to the local file system (via createFsDrain), external logging services (via createAxiomDrain), and various databases (as described in the 'Design calls' section of SKILL.md).
Sanitization: While the skill recommends using auditRedactPreset to filter sensitive PII like passwords and tokens, it does not explicitly address the sanitization of natural language instructions that could be embedded within the logged data fields.

build-audit-logs