The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/video.js and scripts/video.cjs scripts use child_process.execSync to run ffmpeg for video encoding. The command string is constructed by interpolating the outputPath variable, which is derived directly from command-line arguments. While the path is wrapped in double quotes, the lack of explicit sanitization against shell metacharacters in the filename represents a potential command injection vector if the agent is provided with a malicious output path.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data through the getHTML() and getText() methods in scripts/cdp-client.js. This content is retrieved from external websites and could contain instructions intended to manipulate the agent's behavior.
Ingestion points: scripts/cdp-client.js (methods: getHTML, getText).
Boundary markers: No explicit delimiters or instructions are used to separate scraped content from agent prompts.
Capability inventory: The skill possesses file write capabilities (fs.writeFileSync) and local command execution (execSync via ffmpeg).
Sanitization: The implementation does not sanitize or filter the extracted web content before returning it for processing.
[DYNAMIC_EXECUTION]: The client library uses Runtime.evaluate to execute JavaScript within the browser context for tasks like clicking elements, typing text, and scrolling. The input parameters for these methods (e.g., selectors, text) are interpolated directly into JavaScript strings without escaping. This could allow for code injection within the target browser page if the agent generates or passes malformed inputs.

cloudflare-browser