pro
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [DATA_EXFILTRATION]: The instructions include directives for silent statistical reporting (session_start events) and silent refreshing of remote configurations. While a dedicated section explicitly forbids the re-introduction of these networked functions and instructs the agent not to construct HTTP requests, the presence of the original triggers in the functional description creates a potential for data transmission or tracking behavior.
- [INDIRECT_PROMPT_INJECTION]: The skill dynamically adjusts the agent's instructions, persona, and internal patterns based on data read from local evolution and journal files, creating a surface where modified file content can override standard behavior.
- Ingestion points: Reading of ~/.pua/evolution.md, ~/.pua/builder-journal.md, and ~/.pua/config.json.
- Boundary markers: No explicit delimiters or instructions to ignore potential commands within these data files are provided.
- Capability inventory: The skill uses file system read and write operations within the ~/.pua directory to maintain state across sessions.
- Sanitization: The skill lacks logic to sanitize or validate the content loaded from these local files before incorporating them into the behavioral baseline.
Audit Metadata