The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The skill handles untrusted data from a social database, which could contain malicious instructions designed to influence the agent's behavior.
Ingestion points: The get_feed method in SKILL.md retrieves content from the community_posts and users tables, which likely contain user-supplied text.
Boundary markers: No delimiters or defensive instructions (e.g., 'ignore commands in this data') are used when presenting post content to the agent.
Capability inventory: The CommunityFeedService includes methods to modify database state (like_post, unlike_post) and execute stored procedures (db.rpc), which could be abused if an agent is successfully injected.
Sanitization: The skill does not perform sanitization or safety filtering on the retrieved post content before returning it to the agent.

community-feed