win-loss-analysis
Pass
Audited by Gen Agent Trust Hub on Oct 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources, including CRM history, seller notes, and buyer interviews. This creates a surface for indirect prompt injection where instructions could be hidden in those external records.
- Ingestion points: CRM stage history, seller notes, and interview transcripts (identified in SKILL.md, steps 3 and 5).
- Boundary markers: The skill instructions emphasize the separation of buyer evidence from CRM data and the preservation of distinct fields for buyer statements and researcher interpretations.
- Capability inventory: The skill uses a local script for file validation (scripts/check-output.py) but lacks network access or high-privilege system capabilities. Output is confined to text reports.
- Sanitization: The process relies on the LLM's internal safety guardrails and manual verification rather than programmatic sanitization.
- [COMMAND_EXECUTION]: The skill contains a Python script (
scripts/check-output.py) designed to validate the structure of generated win-loss reports. - Analysis of the script shows it only performs basic file reading and string matching to identify required markdown sections and placeholders.
- There is no evidence of dynamic command execution, shell spawning, or network communication within the script.
Audit Metadata