add-eval

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

• Third-party content exposure detected (high risk: 0.85). SKILL.md Step 1A instructs the agent to fetch public web content at runtime (https://docs.convex.dev/llms.txt and then WebFetch of specific Convex docs pages), which is outsider-authored free text that can be ingested into the LLM context.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly instructs the agent at runtime to fetch https://docs.convex.dev/llms.txt and then use retrieved docs pages as ground truth to design prompts/evals, so external content directly influences the agent's instructions.