AgentBoard

The system utilizes dynamic user script injection into the MAIN world of web pages via blob URLs. While efforts are made for sanitization (removing `export`, wrapping in IIFE) and Trusted Types support, direct script injection is an inherent security risk. A malicious user script or a vulnerability in the parsing/injection mechanism could lead to browser context compromise. The `fetch_url` tool can perform CORS-bypassing network requests, optionally including user credentials, which is a powerful capability that could be abused by a malicious LLM prompt or a compromised user script. The `parseMetadataObject` function in `script-parser.ts` and metadata extraction in `src/options/webmcp-scripts.ts` use `new Function()` to parse user-provided metadata, which is a risk if malicious JSON or JavaScript is injected there, though contextually limited. No hardcoded secrets were found. API keys are handled via secure browser storage.