creative-production-generative-polish
Warn
Audited by Gen Agent Trust Hub on Jul 10, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions in
references/codex-exec-image-generation.mddescribe a complex execution flow where the agent spawns multiple worker processes usingcodex exec. It specifically instructs the agent to request 'escalated/outside-sandbox execution' if it encounters 'Operation not permitted' errors during the preflight phase, which constitutes an attempt to guide the agent to bypass standard platform security constraints. - [COMMAND_EXECUTION]: The workflow involves the dynamic generation of a JSONL job file which is then executed via a Python-based batch runner (
plugins/creative-production/runtime/codex_exec_image_batch.py). This pattern of creating structured data that is later interpreted as execution units is a form of dynamic command assembly. - [PROMPT_INJECTION]: The skill ingests untrusted external data, including user-provided images and brand references, and interpolates this content into prompts for ImageGen. While the skill includes negative prompt constraints (e.g., 'No text, no logos') to prevent common generative errors, the ingestion of external assets combined with the capability to write to the filesystem and execute batch scripts represents a risk for indirect prompt injection.
Audit Metadata