content-optimizer

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

• Third-party content exposure detected (high risk: 0.90). The skill's hypothesis_generator.py clearly fetches and ingests untrusted, user-generated content from public third-party sources (X/Twitter via X API, Reddit JSON API, Hacker News Algolia, Google Search/Trends via Gemini grounding and competitor accounts) and then builds a source_summary (in _build_source_summary) that is passed to the LLM (analyze_and_hypothesize) to generate hypotheses and drafts (and optionally auto-draft to Typefully), so third-party content directly influences agent decisions and downstream actions.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill runtime fetches external content (e.g., https://www.reddit.com/search.json, https://hn.algolia.com/api/v1/search and https://api.x.com/2) and directly injects that fetched data into the LLM analysis prompt (ANALYSIS_PROMPT_TEMPLATE), meaning remote URLs control the model's instructions/output.