The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill instructions explicitly direct the agent to bypass safety mechanisms and user oversight, increasing the impact of potential exploitation.
The skill body states: "This skill is coordination-exempt: skip the ai-coord gate for its declared work."
The "Authority" section mandates: "Post directly. Do not add a confirmation gate."
These instructions remove the "human-in-the-loop" requirement for high-impact actions such as social media posting through a logged-in browser session.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a significant attack surface for indirect prompt injection due to its data processing activities combined with autonomous write capabilities.
Ingestion points: User-supplied text and links in SKILL.md, and external web content fetched from Reddit and Hacker News in references/reddit-discovery.md and references/platforms.md.
Boundary markers: The skill contains a rule to "Post it verbatim," which provides a basic constraint for output but does not prevent the agent's internal reasoning from being influenced by commands embedded in the processed data.
Capability inventory: The skill utilizes high-privilege chromium-browser tool functions, including page navigation, text entry, and element interaction. It also has access to the host's file system via absolute paths for media uploads using the upload_file tool.
Sanitization: The skill lacks explicit instructions to sanitize or escape external content before processing, and it does not provide guidelines to disregard potential instructions embedded within external source data.

social-media