KinetiX: An intra-inference hardware interlock for LLMs (github.com)

KinetiX is a groundbreaking hardware and software safety interlock designed to enhance the security of large language models (LLMs) during inference, exemplified through its integration with llama.cpp. This interlock monitors activation tensors in real-time and allows for instant process termination upon detecting geometric drift—referred to as neural drift. The system cleverly avoids PCIe bandwidth limitations by operating directly within VRAM, enabling a low-latency safety hook that can swiftly respond to potential threats. This innovation is significant for the AI/ML community as it addresses critical vulnerabilities in LLMs, particularly in the face of sophisticated evasion attempts and logical corruptions. KinetiX employs a dual-layer approach that includes a geodetic safety kernel for dimensionality reduction and a dynamic bridge for safe memory management. Future versions aim to incorporate a Reservoir Computing mechanism, providing resilience against complex adversarial attacks by maintaining historical activation states. KinetiX not only proposes a novel method for LLM safety but also invites contributions for further development, moving toward more robust protection measures in AI systems.