native-web-search

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

• Third-party content exposure detected (high risk: 0.90). This skill explicitly performs live internet searches and ingests web results (SKILL.md: "search the internet for the requested topic") and the script (search.mjs — runCodexSearch and runAnthropicSearch) passes a web_search tool and builds prompts that require the model to read and summarize third-party web pages and include full canonical URLs, exposing the agent to untrusted public content that could carry indirect prompt injections.