sglang-bisect-ci-regression

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

• Third-party content exposure detected (high risk: 1.00). This skill explicitly fetches and parses public GitHub Actions scheduled run logs and job output (via gh run view / --log) and even downloads repository tarballs from github.com, so it ingests untrusted, user-generated third-party CI/log content that the agent must read and use to decide bisection, reproduction, and next actions.

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

• Potentially malicious external URL detected (high risk: 0.90). The skill contains a runtime remote-install step that fetches and installs code into the target container (curl -L https://github.com/sgl-project/sglang/archive/refs/heads/main.tar.gz | tar xz && pip install -e ...), which downloads and executes remote code during reproduction.