The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a 'Social Listening' workflow in references/listening.md that requires the agent to ingest and analyze content from external sources.
Ingestion points: Untrusted data enters the agent context through the Reddit JSON API, Hacker News (Algolia), Bluesky Public API, and various RSS feeds.
Boundary markers: The instructions do not define clear delimiters or 'ignore embedded instructions' warnings for the external content being processed.
Capability inventory: The agent is tasked with 'scoring' these posts and 'drafting' comments, which could be influenced by malicious instructions hidden in the social media content (e.g., acrostics or hidden text targeting the agent).
Sanitization: There are no documented steps for escaping or validating the external text before the agent processes it for drafting.
[COMMAND_EXECUTION]: The references/listening.md file contains functional shell command 'recipes' using curl, jq, and xmllint to retrieve data from external web services. While these target well-known platforms (Reddit, Hacker News, Bluesky, YouTube) and use standard data processing, they represent an instruction set for executing shell commands with network capabilities.
[DATA_EXFILTRATION]: The skill provides instructions for the agent to navigate and extract data from authenticated LinkedIn and Twitter/X sessions using the dev-browser tool (references/listening.md). While designed for monitoring public posts, this capability creates a risk surface where the agent could be manipulated into accessing or transmitting sensitive account information or private profile data if it encounters malicious input during its automated browsing tasks.

social