audit-skills

Pass

Audited by Gen Agent Trust Hub on Sep 6, 2026

Risk Level: SAFEDATA_EXFILTRATIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill accesses sensitive local files to perform its audit, including chat histories and platform state information.\n
  • Evidence: The scripts/audit-skills.mjs script reads session histories from ~/.codex/sessions, ~/.claude/projects, and ~/.cursor/projects, and queries the SQLite database at ~/.codex/sqlite/state_5.sqlite.\n- [COMMAND_EXECUTION]: The script executes the rg (ripgrep) utility via Bun.spawn to search through session histories.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from session histories to generate an audit report for the agent.\n
  • Ingestion points: The scripts/audit-skills.mjs script scans .jsonl session files across multiple agent platforms.\n
  • Boundary markers: Absent. Data from logs is parsed and summarized into a report without specific delimiters to prevent the agent from misinterpreting embedded text.\n
  • Capability inventory: File system read/write, local command execution (rg), and SQLite database queries.\n
  • Sanitization: Skill names are regex-escaped, but the underlying user and assistant message content in history logs is not sanitized.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 6, 2026, 08:31 PM
Security Audit — agent-trust-hub — audit-skills