roboflow-training-and-evaluation

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

• Third-party content exposure detected (high risk: 0.70). The SKILL.md explicitly instructs the agent to "search community models on Roboflow Universe" and to "Fork from Universe" (community models/datasets), which are public, user-generated resources that the agent would ingest and use to select checkpoints or data—i.e., untrusted third‑party content that can influence training/decisions (see "Universe Model Search" and "Fork from Universe" references).

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly states that clients may fetch MCP fallback resources at runtime using ReadMcpResourceTool for roboflow://skills//... URIs, and those fetched skill pages would directly supply the agent's instruction/content, so this is a runtime external dependency that can control prompts.