karpathy

CRITICAL E005: Suspicious download URL detected in skill instructions.

• Suspicious download URL detected (high risk: 0.80). The GitHub links point to a well-known public repo (karpathy/autoresearch) and are likely low-risk, but https://astral.sh/uv/install.sh is a direct shell installer hosted on a third‑party domain — a high‑risk distribution vector for malware.

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

• Third-party content exposure detected (high risk: 0.85). The required workflow includes the researcher’s WebSearch/WebFetch step (outsider source: public web content fetched at runtime), which can ingest arbitrary free-form text into the LLM context via the researcher agent’s browsing results.

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

• Potentially malicious external URL detected (high risk: 0.90). The ML training template includes explicit setup steps that clone and run remote code—"git clone https://github.com/karpathy/autoresearch.git" and an alternative installer "curl -LsSf https://astral.sh/uv/install.sh | sh"—which fetch and execute external code as a required dependency for the skill.