blockwise-file-authoring

Warn

Audited by Gen Agent Trust Hub on Nov 3, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill documentation suggests generating Python scripts at runtime and saving them to the /tmp/ directory for execution. This pattern is recommended to bypass complex shell quoting requirements but constitutes dynamic code generation and execution.
  • Evidence: Found in SKILL.md under the "Pitfalls" section: "Escribir el contenido con write_file a un .py y ejecutarlo con python3 /tmp/.py".
  • [INDIRECT_PROMPT_INJECTION]: The skill operates on external, potentially untrusted text data such as transcriptions and project documents.
  • Ingestion points: The verification scripts verify_turns.py and verify_prd_roadmap.py read content from files like PRD.md, ROADMAP.md, and raw transcription files.
  • Boundary markers: The skill mandates the use of quoted heredocs ('EOF') when writing to files to prevent shell interpolation, providing a layer of technical isolation, though it does not define specific content-level delimiters for the agent's context.
  • Capability inventory: The skill allows file writing (write_file), file appending (cat >>), and execution of Python scripts and shell utilities (wc, grep).
  • Sanitization: The provided Python scripts use regular expressions to validate document structure (e.g., timestamps, IDs, DAG cycles) but do not perform sanitization of the text content for potential embedded instructions.
  • [COMMAND_EXECUTION]: The skill relies on various shell utilities to manage and verify file integrity during the multi-block writing process.
  • Evidence: Usage of cat, wc, ls, and grep is explicitly instructed in SKILL.md.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 3, 2026, 10:26 AM