The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines an API for rendering web-based compositions from arbitrary URLs, creating a surface where malicious web content could influence the agent's behavior or exploit the rendering environment.
Ingestion points: The render method (SKILL.md) ingests untrusted data through the compositionUrl parameter and the inputProps object, which are injected into the browser context.
Boundary markers: The documentation does not specify the use of delimiters or warnings to prevent the model from obeying instructions embedded within the rendered content.
Capability inventory: The skill's primary capabilities include file system writes (video output, debug traces), network operations (fetching composition URLs), and subprocess execution (FFmpeg and Playwright).
Sanitization: There is no mention of URL validation or data sanitization procedures for the provided input properties.
[DYNAMIC_EXECUTION]: The RendererOptions interface allows specifying custom paths for system binaries, which can lead to the execution of arbitrary code if the paths are influenced by an attacker.
Evidence: The configuration options ffmpegPath and browserConfig.executablePath allow the underlying library to execute binaries from user-defined paths.
[COMMAND_EXECUTION]: The skill's functionality relies on the execution of external tools for media processing.
Evidence: The documentation describes tools like concatenateVideos and rendering strategies that internally spawn FFmpeg and Playwright subprocesses to handle video encoding and browser automation.

helios-renderer