arxiv-research

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

• Third-party content exposure detected (high risk: 0.80). The skill explicitly fetches and reads public, user-submitted arXiv content (via arxiv_search_papers, arxiv_get_paper and arxiv_convert_to_markdown through the arXiv MCP server), meaning the agent ingests untrusted third-party papers that could contain adversarial or instruction-like content.