harness-security-bench

CRITICAL E005: Suspicious download URL detected in skill instructions.

• Suspicious download URL detected (high risk: 0.70). This is a GitHub pull request on an unfamiliar repository that references running an npm package with npx (which downloads and executes arbitrary code) — the URL itself isn't a direct executable download but following its instructions (or running the referenced package) could be risky if the repo/package are untrusted.

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

• Third-party content exposure detected (high risk: 0.85). The required runtime step shells out to npx -y @metaharness/darwin@~0.3.1 ... which fetches and executes an outsider-authored package from a public registry, and the LLM ingests the resulting parsed markdown report (rawMarkdown) and tables into its context.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill runtime invokes "npx -y @metaharness/darwin@~0.3.1 metaharness-darwin security bench", which fetches and executes remote npm package code at runtime (a required dependency), so it executes external code fetched during runtime.