ai-eval-playbook

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

• Third-party content exposure detected (low risk: 0.10). The runtime workflow (“WebFetch the canonical URL”, “?ask= endpoint”, “claude-in-chrome fallback”) ingests first-party web content from eval.playbook.org.ai, but the only explicit outsider-authored free-text path is the user-supplied ?ask=<question> query, which is directly submitted to the page endpoint and could be treated as free text by the agent.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). This skill explicitly requires fetching and using live content at runtime from https://eval.playbook.org.ai/ (including the example ?ask endpoint https://eval.playbook.org.ai/model-behaviour/level-1-module-evaluation/overview.md?ask=what+is+a+golden+dataset and related pages) to control what the agent says, so the external site directly controls prompts and is a required runtime dependency.