The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted plot descriptions and metadata from the external website missav.ws using the agent-browser tool (specifically in scripts/extract-detail.js). These descriptions are subsequently processed for scoring and recommendation ranking in scripts/lib/analyze-description.mjs and scripts/filter-recommendations.mjs. While the current processing is keyword-based, the enriched results are presented to the agent for final presentation to the user, creating a surface where maliciously crafted web content could influence agent behavior. The skill possesses capabilities for local file system writes (to ~/.missav/) and execution of bundled shell scripts, which increases the potential impact of an injection. No explicit boundary markers or sanitization are applied to the external text before it is returned to the agent context.

missav