rewrite-natural-prose

Pass

Audited by Gen Agent Trust Hub on Sep 10, 2026

Risk Level: SAFEPERSISTENCECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PERSISTENCE]: The skill implements a persistent 'rewriting mode' by modifying host instruction files like AGENTS.md (Codex) and CLAUDE.md (Claude). This allows the skill's instructions to influence the agent's behavior across different tasks and sessions within a workspace or client environment.
  • [COMMAND_EXECUTION]: The skill utilizes a local Node.js script (scripts/manage-rewrite-policy.mjs) to manage these configuration files. The agent is instructed to execute this script to check, preview, enable, or disable persistent modes based on user requests.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied prose for refinement, creating an attack surface for instructions embedded in data to influence the agent.
  • Ingestion points: User passages provided for rewriting in SKILL.md.
  • Boundary markers: Present; the skill instructs the agent to treat quoted prose as data and separate editable text from technical content.
  • Capability inventory: Execution of the manage-rewrite-policy.mjs script which can modify persistent agent configuration files.
  • Sanitization: Present; the agent is directed to use structure checks and preserve technical fields (macros, EJS, variable paths, regex) literally without rewriting them.
  • [SAFE]: The file modification script includes robust security controls: it enforces absolute paths, rejects symbolic links, restricts writes to specific host-approved filenames, and implements a multi-step 'preview and approve' workflow using unique operation tokens and mandatory backups to prevent unauthorized or accidental modifications.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 10, 2026, 08:37 AM
Security Audit — agent-trust-hub — rewrite-natural-prose