The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text and image references to construct API requests, presenting a standard surface for indirect prompt injection.
Ingestion points: User-supplied promptText and referenceImages (URLs or data URIs) are ingested via the skill's instructions and processed by the included Python scripts in SKILL.md.
Boundary markers: The skill provides detailed instructions on structured prompt construction and the use of specific tags (e.g., @product_ref) to define reference roles and maintain control over the generated output.
Capability inventory: The provided Python scripts have the capability to perform HTTP requests to the Runway API and external asset hosts, and they can write image artifacts and manifest files to the local file system.
Sanitization: The provided code implements strict validation of output URLs, including checks against a host allowlist and verification that resolved IP addresses are public. Downloaded images are also validated using magic byte signatures (JPEG, PNG, WebP) and pixel dimension limits to prevent decompression bomb attacks or other media-based exploits.
[EXTERNAL_DOWNLOADS]: The skill downloads generated images from external URLs provided by the Runway API. The included implementation mitigates risks associated with external downloads by requiring a host allowlist and performing SSRF-resistant IP verification before connecting.
[COMMAND_EXECUTION]: The skill includes Python scripts intended for execution by the agent to automate the image generation lifecycle. These scripts demonstrate secure coding practices, such as using atomic file operations and enforcing limits on response sizes and pixel counts.

runway-image