Wolf Defender v2 – reducing false positives in a prompt-injection classifier (huggingface.co)

Wolf Defender v2 has been launched, enhancing the capabilities of its multilingual ModernBERT-based classifier designed for detecting prompt injections and jailbreak-like instructions. This version significantly reduces false positives while maintaining high performance across various contexts, including AI agents, chatbots, and automated code workflows. It features a full-size model with a 2,048-token context window and delivers improved metrics, such as increasing the Qualifire F1 score from 94.17% to 95.14% and hard-benign specificity from 81.57% to an impressive 96.23%. Despite a slight drop in clean-validation F1 score, Wolf Defender v2 focuses on better generalization to challenging benign inputs, which is crucial for real-world AI security applications. The release represents a vital advancement for the AI/ML community, as it strengthens defenses against potential prompt injection attacks—a growing concern as AI technology becomes more prevalent. Wolf Defender v2 was trained on a diverse dataset that includes adversarial examples and multilingual sources, improving its robustness against emerging threats. Its architecture allows it to filter untrusted content effectively, positioning it as a critical component in a layered security framework. This makes it particularly significant for organizations that rely on large language models (LLMs) and seek to enhance their security measures while mitigating risks associated with prompt injection vulnerabilities.