local-convex

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

• Potentially malicious external URL detected (high risk: 0.80). The skill starts container images from ghcr.io/get-convex/convex-backend and ghcr.io/get-convex/convex-dashboard at runtime (compose.yaml) which are fetched from an external registry and executed locally, so these external image URLs are runtime dependencies that execute remote code.

HIGH W008: Secret detected in skill content (API keys, tokens, passwords).

• Secret detected (high risk: 1.00). I flagged the literal MinIO CE login because it contains a specific username/password pair ("minioadmin / minioadminPassWD9001") present directly in the document and used to verify CE login during setup — this is a usable credential (not a placeholder). Other values in the doc that might look sensitive were ignored because they are placeholders or references to env files (e.g., AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY are referenced as "<from ...>" and not provided inline, CONVEX_POSTGRES_PASSWORD and INSTANCE_SECRET are mentioned but not shown, and redacted env files are explicitly referenced). Endpoint IPs, ports, and file paths are not secrets.

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 explicitly instructs running privileged commands (sudo -u, sudo sh via SSH), re-running an installer that writes service/config files on the NAS, and otherwise requires elevated access to modify machine state, so it pushes the agent toward state-compromising actions.