flux-2-klein
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes the
@runcomfy/clitool to interact with image generation models. Instructions specify passing user prompts as JSON payloads to prevent shell injection, and the CLI includes logic to verify output URLs against a whitelist of RunComfy domains. - [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied text prompts and reference image URLs, which are passed to the Flux 2 Klein model. The documentation explicitly acknowledges the risk of image-based prompt injection and provides guidance on prompting patterns to mitigate unintended behavior.
- Ingestion points:
promptfield and reference image URLs provided in the JSON input to the CLI. - Boundary markers: User input is encapsulated within a JSON string via the
--inputflag. - Capability inventory: Network communication with
model-api.runcomfy.netand file writes to the directory specified by--output-dir. - Sanitization: Prompt content is treated as a JSON string; the CLI does not perform shell expansion on the data.
Audit Metadata