influencer-sourcing

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

• Third-party content exposure detected (medium risk: 0.30). In scripts/score_creators.py, the runtime path builds a Gemini prompt that includes outsider-authored creator fields (bio, display_name, plus related profile metadata) from output/raw_creators.json (produced by scripts/source_creators.py scraping Instagram/TikTok), so the LLM ingests free text authored by third-party creators without selecting a specific trusted item beyond the batch.

HIGH W008: Secret detected in skill content (API keys, tokens, passwords).

• Secret detected (high risk: 1.00). I looked for high-entropy, literal values that could act as real credentials. The files do not contain API keys (e.g., sk-...), private key blocks, or environment variable values for GEMINI_API_KEY/APIFY_TOKEN. However several avatar_url fields in the scraped JSON include signed CDN URLs with query parameters like refresh_token, x-expires, and x-signature (URL-encoded base64-like strings). Those x-signature values are high-entropy tokens embedded in URLs and are likely signed access tokens that can grant access to resources until they expire. Because they are literal, high-entropy tokens present in the documentation artifacts, I consider them active secrets to flag.

I did NOT flag:

• The environment variable names (APIFY_TOKEN, GEMINI_API_KEY) because no secret values are present.
• Phone numbers and email addresses in bios (not treated as "secrets" per the provided definition).