The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches titles and text snippets from Reddit, which are attacker-controlled external sources, and writes them into PROPOSALS.md. The instructions then encourage the agent to process these proposals to build new skills. This creates a surface where malicious instructions hidden in Reddit posts could influence the agent's behavior.
Ingestion points: radar.py fetches data from the Reddit JSON API (reddit.com/r/{subreddit}/new.json) in the fetch_subreddit and fetch_search functions.
Boundary markers: Absent. The content is written to markdown files without clear delimiters or instructions to the agent to treat the text as untrusted data.
Capability inventory: The skill workflow suggests the agent can execute commands (e.g., create-skill) based on the content of the generated proposals.
Sanitization: The script performs truncation and basic character filtering for slugs, but does not sanitize the primary text content for prompt injection patterns.
[EXTERNAL_DOWNLOADS]: The skill periodically connects to Reddit's public API to retrieve recent posts and comments.
Evidence: radar.py uses urllib.request.urlopen to fetch data from https://www.reddit.com/r/{subreddit}/new.json and search.json using a custom User-Agent.
[COMMAND_EXECUTION]: The skill is intended to be used alongside other tools, instructing the agent to run command-line operations based on the results of the radar scan.
Evidence: SKILL.md provides examples for triggering the radar via python3 radar.py and suggests downstream execution of skill creation tools.

community-skill-radar