content-style

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (medium risk: 0.30). The required workflow reads outsider-authored Reddit comments/posts at runtime via the “browser_navigate to <post-url>.json then parse OP selftext + replies tree” extraction path in references/cross-model-audit-prompt.md, plus additional Reddit comment/threads extraction in references/definitive-guide-reviews-section.md.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.70). The skill explicitly tells authors/agents to use and populate ~/.ssh/config (editing SSH configuration) and to clone/commit/push files to a GitHub mirror (writing to the filesystem/repo), which directs the agent to change machine state and sensitive config files.