gpt-image-2
Pass
Audited by Gen Agent Trust Hub on Oct 11, 2026
Risk Level: SAFEDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates prompt engineering by interpolating user-supplied input into complex JSON-based templates. It has the capability to write the resulting prompts and images to the local filesystem (
garden-gpt-image-2/) and perform network operations to send these prompts to external APIs. While the templates provide a structural boundary, the skill lacks rigorous input sanitization beyond basic trimming and slugification. - Ingestion points: Command-line arguments in
scripts/generate.jsandscripts/edit.js(e.g.,--prompt). - Boundary markers: Structured JSON templates and explicit argument placeholders (e.g.,
{argument ...}) are used. - Capability inventory: Local file writing (
node:fs/promises) and network requests (fetch). - Sanitization: Input strings are trimmed and filenames are normalized using a slugify function.
- [DATA_EXFILTRATION]: The configuration utility in
scripts/shared.jssearches for API keys and endpoint settings in several local paths, including.envand~/.gateway.env. While this is standard practice for local AI tools to manage credentials, it involves accessing files outside the project directory. - [EXTERNAL_DOWNLOADS]: The skill documentation references images hosted on
cdn.jsdelivr.netand links to a showcase website onmmh1.top. The scripts communicate with the OpenAI API (defaultapi.openai.com) and include functionality to download generated images from URLs returned by the configured API provider.
Audit Metadata