The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides shell commands to interact with the system's infrastructure to retrieve the data required for the audit.
Evidence: Utilizes docker compose exec to execute bash scripts and database queries within the api and postgres containers.
Evidence: Employs bash pseudo-device files (/dev/tcp/qdrant/6333) to perform network operations against internal services.
[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves processing external data (the corpus) which could potentially contain malicious or manipulative instructions.
Ingestion points: Untrusted data is ingested from the Qdrant chunk payload (written to chunks.json) and the Postgres audit_log table (payload_json field) via shell commands.
Boundary markers: The skill does not explicitly define delimiters or specific instructions for the agent to ignore or isolate instructions that might be embedded within the retrieved corpus text.
Capability inventory: The skill relies on shell execution (docker compose exec, tail, grep), internal network access, database access, and file writing capabilities.
Sanitization: The skill instructions do not specify any sanitization, filtering, or validation steps for the content retrieved from the database or vector store before the agent processes it.

answer-audit