The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a 'Social Listening' feature that fetches data from external, untrusted sources including Reddit, Hacker News, and Bluesky. This content is then processed by the agent to generate scores and draft comments, creating a vulnerability to indirect prompt injection where malicious instructions embedded in social media posts could manipulate the agent's behavior.
Ingestion points: The file references/listening.md defines workflows for fetching data from https://www.reddit.com, https://hn.algolia.com, and https://public.api.bsky.app.
Boundary markers: The skill partially mitigates this risk by instructing the agent to 'Always ask before posting — output drafts, user approves and posts manually,' ensuring a human-in-the-loop review.
Capability inventory: The skill utilizes curl for network requests and dev-browser for navigating authenticated social media feeds.
Sanitization: There is no evidence of sanitization or escaping of the external content before it is interpolated into the agent's context.
[COMMAND_EXECUTION]: The file references/listening.md provides several shell 'recipes' using curl and jq for data collection. These recipes utilize placeholders like KEYWORD and CHANNEL_ID which are intended to be replaced at runtime. If the agent populates these placeholders with unsanitized user input, it could lead to command injection within the shell environment where the agent executes these tools.

social