image-style-transfer

Pass

Audited by Gen Agent Trust Hub on Jun 22, 2026

Risk Level: SAFECOMMAND_EXECUTIONCREDENTIALS_UNSAFEDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes the vendor-provided 'meitu' CLI tool to perform various image stylization tasks based on user input.\n- [CREDENTIALS_UNSAFE]: The skill is designed to read authentication keys from the '/.meitu/credentials.json' file to authenticate its requests to the Meitu OpenAPI.\n- [DATA_EXFILTRATION]: The skill requires access to read configuration files and registry metadata from the user's home directory ('/.meitu/') and local workspace directories to correctly route image processing requests.\n- [PROMPT_INJECTION]: The skill interpolates user-provided styling descriptions into a shell command, creating a surface for indirect prompt injection.\n
  • Ingestion points: User-provided style descriptions (prompt) and image URLs are ingested via the CLI command arguments.\n
  • Boundary markers: The 'prompt' value is wrapped in double quotes within the shell command template, but no additional boundary markers or instructions to ignore embedded commands are present.\n
  • Capability inventory: The skill can execute the 'meitu' binary and has filesystem write permissions to the './output/' and workspace directories.\n
  • Sanitization: There is no evidence of sanitization or escaping of the user-provided prompt string before it is passed to the shell environment for execution.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 22, 2026, 04:39 AM
Security Audit — agent-trust-hub — image-style-transfer