kai-hook-bench
Pass
Audited by Gen Agent Trust Hub on Sep 28, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The skill executes multiple internal Python scripts to perform core tasks: data collection (
scripts.audit.collect), brand analysis (brand_pulse.py), quality scoring (four_us_score.py), and memory management (lesson_capture.py). These scripts run within the shell environment provided to the agent. - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from untrusted external sources, which could lead to accidental or malicious instruction execution if the data contains prompt injection strings.
- Ingestion points: The skill mines data from Reddit digests, customer reviews, and external websites via the
scripts.audit.collecttool to populateMARKETING.mdandpain-inputs.md. - Boundary markers: There are no explicit delimiters or instructions provided to the agent to disregard instructions potentially embedded within the ingested marketing data.
- Capability inventory: The agent has shell execution capabilities (running Python scripts) across all phases of the workflow.
- Sanitization: The skill description does not specify any sanitization, filtering, or validation steps for the data retrieved from external URLs or social media mining.
- [EXTERNAL_DOWNLOADS]: Uses the
scripts.audit.collectmodule to fetch content from user-provided URLs for market research and hook generation.
Audit Metadata