ljg-skill-map

Pass

Audited by Gen Agent Trust Hub on Mar 11, 2026

Risk Level: SAFECOMMAND_EXECUTIONPROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [COMMAND_EXECUTION]: Executes a local bash script scripts/scan.sh to iterate through directories and extract YAML frontmatter from files. While functional for a skill map, it represents a direct shell execution capability.\n- [DATA_EXFILTRATION]: Performs recursive reads within the ~/.claude/skills/ directory. Although no network exfiltration is observed, the skill exposes the documentation and configuration of all installed skills to the active context.\n- [PROMPT_INJECTION]: Vulnerable to indirect prompt injection (Category 8). Maliciously crafted descriptions in other installed skills could attempt to override agent instructions when processed during the categorization and ASCII rendering phases. \n
  • Ingestion points: Metadata extraction from SKILL.md files via scripts/scan.sh. \n
  • Boundary markers: Absent; descriptions are interpolated into the prompt for classification without delimitation or instructions to ignore embedded commands. \n
  • Capability inventory: Local file system read and bash script execution. \n
  • Sanitization: Includes shell-level escaping for quotes and backslashes in the script, but lacks natural language sanitization for the LLM processing stage.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 11, 2026, 07:22 AM