web-search

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

• Third-party content exposure detected (high risk: 0.85). Outsider free text is ingested into the agent’s LLM context via model-mediated providers (e.g., openai/anthropic/xai): the skill sends the user query to the provider’s web_search tool and then normalizes/returns cited_text/text excerpts (from third-party web pages) into combined.json, which the agent commonly pipes into its LLM for summarization.