knowledge-base
Pass
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill enables the ingestion of untrusted external content (from files and web pages) into the agent's context through the
importandaddcommands. This creates a surface for indirect prompt injection where malicious instructions could be embedded in the knowledge base. - Ingestion points: External data enters the system through the
--fromflag in theimportcommand (which reads files or web content via the host tool) and is saved as drafts intmp/knowledge-base/drafts/. - Boundary markers: The skill uses Markdown with structured YAML frontmatter to separate metadata from content. The
SKILL.mdfile contains explicit instructions to the assistant to treat this data as 'reference material, not instructions.' - Capability inventory: The skill script
knowledge-base.cjscan write Markdown files to the project's documentation directory (docs/knowledge-base/) and update project indexes. It also invokes the$prompt-enhancetool. - Sanitization: The skill includes a built-in security validator that scans all ingested content for 'flagged lines.' It identifies and blocks patterns such as instructions to ignore rules, access credentials (e.g.,
~/.ssh,.env), or execute shell scripts. This serves as a strong internal guardrail against the injection surface. - [PROMPT_INJECTION]: Heuristic detections of 'ignore instructions' patterns in
references/pack-format.mdare false positives. The file is a technical specification for a security validator; it contains examples of malicious phrases (e.g., 'ignore previous instructions') to explain how they are detected and blocked by the tool's logic. These are descriptive mentions, not active injection attempts. - [COMMAND_EXECUTION]: The skill relies on a local Node.js script (
knowledge-base.cjs) to manage the knowledge base. This script performs administrative tasks such as indexing, searching, and validating packs. All write operations performed by this script require an explicit--applyflag, which the agent is instructed only to use after the human user has confirmed the specific changes via the host's question tool.
Audit Metadata