gpt-atelier
Pass
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill includes functionality to execute a shell command to open a generated HTML comparison report in the user's default browser.
- Evidence:
subprocess.run(["open", str(html_path)], check=False)inscripts/compare_models.py. - [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection because it ingests untrusted user content and reference images, which are then processed by AI models with the ability to influence local file operations.
- Ingestion points: User-provided prompts and image paths are accepted in
scripts/generate_image.py,scripts/edit_image.py,scripts/compose_images.py, andscripts/converse_image.py. - Boundary markers: Absent; the skill interpolates user inputs directly into the API request payloads without explicit delimiters or warnings to the model to ignore embedded instructions.
- Capability inventory: The skill performs network operations via the
openailibrary, writes image files to the local./outputdirectory via thesave_imagesfunction inscripts/openai_images.py, and executes local commands viasubprocess.runinscripts/compare_models.py. - Sanitization: Absent; no validation, escaping, or filtering is applied to the user-supplied prompts or image metadata before they are sent to the OpenAI API or used in filename construction.
Audit Metadata