gpt-image-2

Pass

Audited by Gen Agent Trust Hub on Jul 13, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTIONCREDENTIALS_UNSAFE
Full Analysis
  • [COMMAND_EXECUTION]: The skill invokes the runcomfy CLI for model operations. It specifies that user input is passed as a JSON string to mitigate shell injection risks.
  • [EXTERNAL_DOWNLOADS]: Requires the installation of @runcomfy/cli from the official NPM registry and fetches the skill from the agentspace-so/runcomfy-skills GitHub repository.
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection through its ingestion of user-provided prompts and external image URLs. The documentation specifically warns about image-based prompt injection risks. All Category 8 requirements are met: ingestion points are identified in the prompt and images fields; boundary markers are noted at the CLI level but absent at the prompt level; capabilities include CLI execution and file writes; sanitization is performed by the CLI to prevent shell injection.
  • [CREDENTIALS_UNSAFE]: The skill manages authentication tokens by storing them in ~/.config/runcomfy/token.json or via the RUNCOMFY_TOKEN environment variable. It correctly advises setting restrictive file permissions (mode 0600) for local token storage.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 13, 2026, 04:10 AM
Security Audit — agent-trust-hub — gpt-image-2