HiveTraceGuard-Pro: A Compact Generative Guardrail for Prompt Injection, Jailbreaks, and Adversarial Obfuscation
HiveTraceGuard-Pro is a 0.6B generative guardrail LoRA-tuned from Qwen3-0.6B, trained on Russian and English, using a binary safe/unsafe scoring rule for the final target turn. Its training corpus pairs harmful examples with benign examples from the same domain and applies eight obfuscation transforms to both labels. In a comparison with thirty-four other guards on nineteen benchmark groups (sixteen public), its aggregate key is 0.7432, behind 0.7641 and 0.7552 for the two higher-scoring guards. Over the sixteen public groups alone, its key is 0.7153 and four of the thirty-four other suite guards score higher. In a fifteen-model comparison, HiveTraceGuard-Pro has the highest clean Russian robustness combined-F1 (0.88) and Russian prompt-injection recall (0.999).
HiveTraceGuard-Pro is a compact 0.6B generative guardrail model LoRA-tuned from Qwen3-0.6B, designed to detect prompt injection, jailbreaks, and adversarial obfuscation in Russian and English. It uses a binary safe/unsafe scoring rule for the final target turn and is trained on paired harmful and benign examples with eight obfuscation transforms. In benchmark comparisons, it achieves an aggregate key of 0.7432 across nineteen benchmark groups, ranking behind two higher-scoring guards, and 0.7153 on sixteen public groups where four other guards score higher. It demonstrates strong Russian robustness with a combined-F1 of 0.88 and prompt-injection recall of 0.999 in a fifteen-model comparison.
The model's 0.6B parameter size and LoRA tuning from Qwen3-0.6B indicate a focus on efficient, deployable guardrails. The use of paired harmful/benign examples and eight obfuscation transforms suggests robustness to adversarial input variations. The binary scoring rule for the final target turn simplifies decision-making but may limit nuanced risk assessment. Performance metrics show competitive but not leading aggregate scores, with notable strength in Russian prompt-injection recall, indicating specialized training data for Russian language threats.
The development of compact, multilingual guardrail models addresses the growing need for safety mechanisms in production LLMs, especially for non-English languages. The benchmark results highlight a competitive landscape with multiple guardrail solutions, where no single model dominates all metrics. The focus on Russian suggests a response to regional security requirements or underserved language coverage in AI safety.
HiveTraceGuard-Pro offers a lightweight, multilingual guardrail solution that can be deployed with lower computational overhead compared to larger models. Its strong Russian performance addresses a specific market need, potentially enabling safer AI deployment in Russian-speaking regions. However, its aggregate performance behind some competitors may limit adoption in environments requiring top-tier safety across all benchmarks.
Observable next signals include further improvements in aggregate benchmark scores, expansion to additional languages, and integration into commercial LLM safety stacks. The gap between HiveTraceGuard-Pro and higher-scoring guards may drive research into better training data or architecture choices. Adoption in Russian-speaking markets could be a key indicator of real-world effectiveness.