yetong-promo
Pass
Audited by Gen Agent Trust Hub on Oct 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from user-provided profile files (e.g.,
person-dna.md) and product descriptions to generate personalized marketing content. - Ingestion points: External markdown files and user-supplied product data cards (
references/product-card.md). - Boundary markers: Instructions explicitly direct the agent to read only specific fields and avoid guessing or inferring unconfirmed facts. It uses delimiters and structured references for data.
- Capability inventory: The skill uses a subprocess call to a local Python script and performs file read operations.
- Sanitization: A dedicated script (
scripts/check_profile.py) performs structural and schema validation on the profile file before it is used for generation. - [COMMAND_EXECUTION]: The skill prompts the agent to run a local script included in the skill package for data validation purposes.
- Evidence:
python3 scripts/check_profile.py <用户明确指定的档案路径>inSKILL.md. - Analysis: The execution is limited to a specific script provided with the skill. The user provides the file path argument, which is then processed by the script's internal logic.
Audit Metadata