The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/image.sh performs unsafe URL construction by directly appending shell variables like $MODEL, $WIDTH, and $HEIGHT to the query parameter string without performing URL encoding. This oversight allows a user or automated process to inject additional, unintended query parameters into the API request.
[DYNAMIC_EXECUTION]: The script scripts/tts.sh is vulnerable to JSON injection. The shell variables $VOICE and $FORMAT are interpolated directly into the jq filter string used to build the request body, rather than being passed as safe arguments via --arg. This enables the injection of unauthorized fields or the creation of malformed JSON payloads.
[EXTERNAL_DOWNLOADS]: The skill makes network requests to gen.pollinations.ai to interface with AI models and downloads generated media files (images, videos, and audio) to the local filesystem using curl.
[INDIRECT_PROMPT_INJECTION]: The skill accepts untrusted data which could contain malicious instructions designed to influence the behavior of the integrated AI models.
Ingestion points: User-supplied text prompts in scripts/chat.sh, scripts/image.sh, and scripts/tts.sh, as well as external resource URLs via the --image-url parameter in scripts/image.sh.
Boundary markers: There are no delimiters or specific instructions to the models to ignore embedded commands within the processed data.
Capability inventory: The skill utilizes curl for network communication and has the capability to write files to the local system across all provided scripts.
Sanitization: No validation or sanitization of the input content is performed beyond a simple character substitution in scripts/image.sh.

pollinations