fal-ai-image
Pass
Audited by Gen Agent Trust Hub on Sep 14, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input via the
--promptand--image-urlsarguments, which are interpolated into JSON payloads for external API calls. - Ingestion points: Prompt and image URL parameters in
generate.shandedit.sh. - Boundary markers: The skill uses
json_quoteto escape strings but does not employ specific delimiters for the target model to prevent instruction override within the prompt itself. - Capability inventory: The agent can perform network requests (
curl) and write files to the local system. - Sanitization: Input data is escaped using Python's
json.dumpsto ensure valid formatting in the outgoing request. - [DYNAMIC_EXECUTION]: Utility functions in
scripts/common.shutilize static Python code snippets via heredocs to perform data processing tasks like JSON quoting, CSV parsing, and dimension validation. This is a common pattern used to extend the capabilities of shell scripts. - [EXTERNAL_DOWNLOADS]: The skill downloads generated images from fal.ai storage to the local filesystem using
curl. These downloads target the official API infrastructure of the service provider.
Audit Metadata