The Canary's Echo: Auditing Privacy Risks of LLM-Generated Synthetic Text