Tenable Enhances AI Security Detectors with Native Multi-Language Support
Tenable has upgraded its AI security detectors to natively support nine additional languages, moving beyond English-only analysis to improve threat detection accuracy and speed.
Tenable has significantly advanced its AI security capabilities by enabling its threat detectors to natively understand and process threats in nine languages beyond English. This crucial enhancement addresses the limitations of English-centric security tools, which previously required translating non-English threats into English, a process that was slow, costly, and prone to losing critical context.
The company developed an innovative agentic translation pipeline to achieve this. This pipeline focuses on generating high-fidelity, label-preserving training and benchmark data across the target languages. By ensuring that the security-relevant meaning and the original threat label are maintained through the translation process, Tenable can build a single, robust multilingual model.
This new approach contrasts sharply with the previous "translate-then-detect" method. The old process introduced significant latency, as large translation models needed hundreds of milliseconds to process even short inputs. It also incurred higher costs due to the per-request translation pass and added operational overhead for managing an additional heavyweight model. Furthermore, the translation process itself could degrade the tool's effectiveness by distorting the original intent and meaning of threats.
To overcome these challenges, Tenable's agentic pipeline employs a three-stage process: "Translate," "Estimate," and "Gate." In the "Translate" stage, a frontier LLM performs "transcreation," a fine-tuned translation that prioritizes preserving security context. The "Estimate" stage involves a panel of independent judges scoring each translation for fluency, meaning integrity, and label preservation. Only translations that pass all quality checks in the "Gate" stage are retained.
Two critical recovery loops are integrated to handle edge cases. "Refusal handling" ensures that when a primary translator declines to process sensitive content, a controlled open-source model takes over, with its output still subject to the quality gate. The "Refinement" loop sends translations that fail the quality gate back for another attempt, preventing data loss. This pipeline is designed to be adaptable, capable of serving any future detection task requiring multilingual support.
The core of Tenable's solution is a pretrained multilingual model that was fine-tuned for this specific application. This strategy leverages the model's existing cross-lingual understanding, requiring less data per language and yielding faster, more accurate results. The result is a single model that can detect threats natively in all 10 supported languages, offering a substantial improvement in speed—estimated to be 20 to 800 times faster per item—and a significant reduction in cost compared to the previous translation-dependent method.
This move towards native multilingual support is vital for comprehensive AI security. Prompt attacks, malicious content, and policy violations can originate in any language, and an English-only detector leaves organizations vulnerable to threats they cannot even recognize. By supporting a wider array of languages, Tenable aims to provide more robust and equitable protection against the evolving landscape of AI-driven threats.