The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute a shell command using curl to verify YouTube thumbnail availability. This creates a command injection vulnerability because the VIDEO_ID variable, extracted from user-supplied URLs, is interpolated directly into the command string (curl -s -o /dev/null -w "%{http_code}" URL). An attacker could provide a crafted URL that causes the agent to execute arbitrary shell commands via subshell or command chaining syntax.
[INDIRECT_PROMPT_INJECTION]: The skill fetches and processes content from external, untrusted sources including recipe websites and YouTube video descriptions. It lacks explicit instructions to treat this content as data rather than instructions, potentially allowing a malicious recipe to hijack the agent's behavior.
Ingestion points: External recipe URLs and YouTube video descriptions (defined in the YouTube watch URLs section of SKILL.md).
Boundary markers: None present to separate external content from agent instructions.
Capability inventory: Command execution (curl), file write operations (overwriting files in the vault), and external network requests.
Sanitization: The skill applies formatting constraints (plain text, specific capitalization, sub-listing), which provides limited structural filtering but does not prevent malicious instruction following.
[EXTERNAL_DOWNLOADS]: The skill utilizes tools like WebFetch and Playwright to retrieve data from arbitrary external domains not limited to a whitelist. While the YouTube domain is recognized as a well-known service, the skill also targets unknown third-party blog posts and recipe pages referenced in video descriptions.

recipe-cleanup