How Tenable brought multi-language support to security detectors with agentic translation

Tenable One AI Exposure’s threat detectors were initially English-centric, relying on a translate-then-detect process that translated non-English inputs into English before running detection. This method caused latency, high costs, operational overhead, and reduced detection accuracy due to loss of meaning in translation. To improve this, Tenable built an agentic translation pipeline that uses a frontier large language model (LLM) to produce transcreated translations that preserve the original label and security context. Each translation is quality-checked by multiple judges and refined or replaced by fallback models if necessary. This pipeline generates high-quality multilingual training and test data across nine additional languages, enabling a single multilingual model to detect threats natively in all supported languages. This approach eliminates the need for online translation during detection, significantly improving speed and cost efficiency while maintaining detection fidelity.

The impact is an enhanced ability to detect AI-related threats, harmful content, and policy violations in multiple languages with higher accuracy and much faster processing times compared to previous English-only detection methods that relied on translation. This reduces the risk of missing or misclassifying threats in non-English inputs, thereby improving overall security posture against multilingual threat actors. There is no indication of a vulnerability or exploit; rather, this is an advancement in detection technology.

This content describes an improvement in AI threat detection technology rather than a vulnerability requiring mitigation. No patch or remediation is applicable. Organizations should consider adopting or integrating multilingual detection capabilities to improve coverage of non-English threats. No urgent action is required as this is a product enhancement.