suggesting-path-cleaning-rules

Pass

Audited by Gen Agent Trust Hub on Sep 6, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data (sampled website pathnames) which is then passed to an LLM to generate configuration rules. This architecture is susceptible to indirect prompt injection if a malicious user creates pathnames designed to manipulate the LLM's logic.
  • Ingestion points: The sample_pathnames function in service.py fetches raw $pathname values from web analytics traffic.
  • Boundary markers: The documentation does not specify if the prompt templates use delimiters or explicit instructions to prevent the LLM from executing commands found within the sampled pathnames.
  • Capability inventory: The skill utilizes an LLM gateway, performs regex compilation via the re2 engine, and includes tools to modify project-level path_cleaning_filters.
  • Sanitization: Rules are validated for syntax and matches using re2 before storage, but the skill does not mention sanitizing the input pathnames to filter out potential injection strings before they reach the LLM.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 6, 2026, 07:58 AM
Security Audit — agent-trust-hub — suggesting-path-cleaning-rules