benchmark

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

• Third-party content exposure detected (high risk: 0.70). This skill navigates to arbitrary external URLs and runs the browse daemon ($B goto and $B eval "performance.getEntriesByType(...)") to ingest performance and resource data from public third‑party websites (the supplied to /benchmark and auto-discovered pages in Phase 2/Phase 3), which the agent reads and uses to make decisions about regressions and recommendations.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill's setup step can run a runtime command that pipes a remote installer to the shell ("curl -fsSL https://bun.sh/install | bash"), which fetches and executes remote code and is required for the one-time build, so it is a high-risk external dependency.