kb-harvest

Warn

Audited by Gen Agent Trust Hub on Apr 25, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill allows the agent to read content from "arbitrary local directories" and "sibling repos." While it excludes common folders like .git/ and node_modules/, it does not explicitly prevent access to sensitive user configuration files such as ~/.ssh/ or .env files. If a user or malicious instruction provides a path to these sensitive files, the skill would "harvest" and "distill" their contents into the project's knowledge base, potentially exposing credentials.- [EXTERNAL_DOWNLOADS]: The skill utilizes a tool to fetch and retrieve content from arbitrary, user-provided web URLs.- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests untrusted data from external sources and instructs the agent to convert that data into "actionable rules in imperative voice."
  • Ingestion points: External web URLs (retrieved via tool) and local/sibling repository files (retrieved via file tools).
  • Boundary markers: The skill incorporates a human-in-the-loop verification step using a confirmation prompt before writing any content to the disk.
  • Capability inventory: The skill has read access to the local file system and web, and write access to the project's documentation files.
  • Sanitization: The skill includes a rule to strip authentication tokens from fetched URLs, but lacks logic to sanitize the content for malicious instructional patterns embedded in the source text.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 25, 2026, 01:21 AM