The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill represents an indirect prompt injection surface because it ingests untrusted data from a public API and uses it to generate content that is automatically published.
Ingestion points: RealPolymarketTool fetches event titles and market data from https://gamma-api.polymarket.com/events.
Boundary markers: None. The _generate_script method in agent.py interpolates the market title directly into the script template without sanitization or instructions to ignore embedded commands.
Capability inventory: The skill possesses the youtube_post capability (via RealYouTubeTool), which allows it to write data to an external service (YouTube).
Sanitization: No escaping or filtering is applied to the market data before it is processed by the LLM or uploaded to YouTube.
[DATA_EXPOSURE]: The skill handles several sensitive credential types and file paths.
The system is designed to manage client_secrets.json and token.json for YouTube OAuth 2.0 access.
The .env.example file highlights that the skill expects highly sensitive variables such as PRIVATE_KEY (Web3), YOUTUBE_CLIENT_SECRET, and various AI provider API keys.
Access to these files is necessary for the skill's functionality but presents a risk if the environment is compromised.
[COMMAND_EXECUTION]: The RealYouTubeTool in tools.py utilizes flow.run_local_server(port=0). This command spawns a local web server to handle the OAuth 2.0 authorization code flow. While standard for desktop applications, this behavior might be unexpected in a headless or restricted agent execution environment.

fastkol-web3-content-pipeline