<p>As large language models (LLMs) become integral to modern software applications, ensuring their secure deployment on untrusted infrastructure has become critical. When LLMs are deployed on untrusted GPUs, malicious hosts can exfiltrate proprietary model weights or execute unauthorized inferences, threatening intellectual property. We introduce TransLock, a self-locking watermark framework that enables secure LLM deployment in real-world software environments while maintaining practical performance. First, it strategically partitions model parameters, maintaining security-critical weights in protected CPU memory or Trusted Execution Environments while leveraging untrusted GPUs for computational efficiency. Second, ephemeral watermarking continuously refreshes these protected parameters using short-lived random seeds, ensuring that any cached or stolen model components rapidly become obsolete. Third, a puzzle-based access control mechanism enforces sequential computational costs before each inference session, throttling brute-force extraction attempts by preventing parallel speedups. TransLock’s implementation on LLaMA-2 preserves performance with minimal overhead while preventing untrusted GPUs from reconstructing usable models or producing valid outputs.</p>