china-3c-marketing-copilot
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from CSV files and user-generated campaign text, interpolating them directly into instructions for persona simulation and quality checks.
- Ingestion points: Data enters the system via
scripts/preprocess.py(readingraw_data.csv) and thesimulate_comment_sectionfunction (processingcampaign_text). - Boundary markers: The prompt construction examples use simple string interpolation (f-strings) without delimiters or instructions to ignore embedded commands (e.g.,
f"As a {persona_type}... critique: {campaign_text}"). - Capability inventory: The skill possesses file system access, including writing capability in
preprocess.py(df.to_json) and reading across theknowledge-base/directory. - Sanitization: The preprocessing script performs basic schema validation but does not sanitize input text for malicious prompt injection patterns.
- [COMMAND_EXECUTION]: The skill instructs users to execute shell commands for installation and data processing.
- Evidence: The documentation specifies usage of
git clone,pip install, and running a local Python script (python scripts/preprocess.py).
Audit Metadata