Shield: A 118M model for detecting prompt injections and jailbreaks (huggingface.co)

The newly announced Shield Small is a 118 million-parameter AI model designed for detecting prompt injections and jailbreaks in untrusted text inputs. With compact ONNX weights of just 118 MB, it operates efficiently on local CPUs, enabling applications to pre-screen various text formats—such as documents, web pages, and messages—before they reach AI agents. The model outputs an injection score alongside a binary verdict, allowing developers to implement custom blocking or review mechanisms based on the provided assessment. This release is significant for the AI/ML community as it directly addresses the rising concerns around the security and robustness of AI systems, particularly regarding nefarious attempts to manipulate AI through prompt injections. Shield Small was fine-tuned using a diverse dataset that includes benign texts and labeled attack examples, achieving a mean category-balanced accuracy of 84.3%. Its ability to run offline ensures data privacy for applications while providing necessary safeguards against malicious inputs. The model showcases the balance between effectiveness and efficiency, underlining the importance of integrating AI security measures in real-world applications.