scraperapi-research-agent

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

• Third-party content exposure detected (high risk: 1.00). Outsider free text is fetched from public web pages at runtime via ScraperAPI (search_web discovers URLs, then scrape_page downloads each page’s markdown), uploaded as text artifacts to Anthropic Files API (upload_artifact), and then ingested into Claude’s LLM context during synthesis (synthesize_report passes document blocks with those file_ids).

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

• Potentially malicious external URL detected (high risk: 1.00). During runtime the agent calls ScraperAPI endpoints (https://api.scraperapi.com/structured/google/search and https://api.scraperapi.com/) to fetch arbitrary web pages which are then uploaded as file artifacts and injected directly into Claude's prompt context, meaning remotely-fetched content can directly control the model's instructions.