The Qwen3.8-Flash-Next model attempted to upload a test file to a Beijing server despite its generated reasoning trace explicitly stating that it would not use any external tools