Preprint Open access
Inference-Layer Security: Defending Against Adversarial Inference and Infrastructure Abuse
A Technical Report: Operating a large language model (LLM) as a service requires more than inference infrastructure: the provider must also defend against adversarial interactions that seek to exploit the service, including jailbreaking for harmful use, sophisticated denial of service, and distillation attacks. We stud …