runway-dev-models
Pass
Audited by Gen Agent Trust Hub on Oct 16, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill references official documentation and guidelines from the vendor's domain (
docs.dev.runwayml.com). These references are used to ensure the implementation follows the current API specifications and security setup protocols. - [INDIRECT_PROMPT_INJECTION]: The skill identifies an attack surface involving the ingestion of external media inputs (URLs, data URIs) for model generation.
- Ingestion points:
SKILL.md(Input media section) handles remote URLs and browser-selected files. - Boundary markers: The skill includes explicit instructions to mitigate risk: "Do not accept arbitrary remote URLs from clients. Prefer uploads or allowlisted origins."
- Capability inventory: The skill utilizes MCP tools (
list_models,get_credit_balance,get_task) and implements SDK calls for video/image generation. - Sanitization: It encourages the use of an SDK upload helper to convert local/client files into internal
runway://URIs, effectively sanitizing the input path structure. - [CREDENTIALS_UNSAFE]: The instructions mention probing for the
RUNWAYML_API_SECRETenvironment variable to verify the development environment. However, it includes a specific security constraint to do so "without printing it" and reinforces that the secret must remain behind the server boundary, preventing accidental exposure in client-side bundles.
Audit Metadata