learn

Pass

Audited by Gen Agent Trust Hub on May 16, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Uses shell commands such as find, wc, grep, sed, and xargs in SKILL.md (Step 1) to identify existing assistant configuration files and skill definitions within the local file system.
  • [PROMPT_INJECTION]: Vulnerable to indirect prompt injection because it processes untrusted data from conversation history to generate instructions that are persisted in project-level configuration files (e.g., CLAUDE.md, .cursorrules, .continuerc.json).
  • Ingestion points: The skill analyzes conversation history as defined in SKILL.md (Step 2) to extract "Discoveries", "Workflows", and "Corrections".
  • Boundary markers: Lacks programmatic delimiters to isolate untrusted data during analysis, but utilizes a procedural boundary via Step 5's mandatory human review process.
  • Capability inventory: The skill has permissions to write to multiple assistant configuration files and create new executable skill definitions in the skills/ directory (Step 6).
  • Sanitization: Incorporates a "Minimum viable rule text" audit to trim rule content and requires explicit user approval before any file system modifications are executed.
Audit Metadata
Risk Level
SAFE
Analyzed
May 16, 2026, 11:21 PM
Security Audit — agent-trust-hub — learn