pythia-conventions
Pass
Audited by Gen Agent Trust Hub on Aug 28, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes database object names (external data) to derive coding patterns, which are then stored in project configuration files. This creates a surface where maliciously named database objects could influence the agent's behavior in subsequent sessions.
- Ingestion points: Database object names are ingested through the
pythia conventions --scancommand as described in SKILL.md. - Boundary markers: The instructions include a manual reconciliation step in the workflow where the agent is directed to ask the developer about discrepancies, providing a human-in-the-loop check.
- Capability inventory: The skill has the capability to write local project files, specifically
.pythia/conventions.jsonandconventions.md(SKILL.md). - Sanitization: There are no instructions for automated sanitization or validation of the object names before they are tokenized and saved into the convention files.
Audit Metadata