run-batch-codex-research

Pass

Audited by Gen Agent Trust Hub on May 19, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is designed to process untrusted external data from sources such as inputs.txt, urls.txt, or CSV files, which creates an indirect prompt injection surface.
  • Ingestion points: Untrusted data enters the context via the input resolution process described in Step 1, where it reads from files or pasted lists.
  • Boundary markers: The skill uses a structured work directory layout to separate frozen templates (template.md), rendered prompts (prompts/), and final answers (answers/), creating clear boundaries between stages of execution.
  • Capability inventory: The skill has the capability to execute external binaries on the system path (e.g., codex, claude, gemini, ollama) and perform file system writes and renames (Step 5).
  • Sanitization: The skill explicitly mitigates command injection risks by mandating in-process prompt rendering (Rule 3) and forbidding shell string concatenation for input interpolation (Step 3).
  • [COMMAND_EXECUTION]: The skill provides templates for executing system commands and managing a pool of worker processes.
  • Evidence: The skill defines shell command shapes using timeout, xargs, and redirection patterns for managing LLM CLI execution and log capture (Step 4, Step 5, and Single-CLI command shape).
Audit Metadata
Risk Level
SAFE
Analyzed
May 19, 2026, 02:36 PM
Security Audit — agent-trust-hub — run-batch-codex-research