Tenable details native multilingual AI threat detection in 10 languages for Tenable One AI Exposure
Tenable describes how it moved Tenable One AI Exposure detectors (jailbreaks, prompt injection, harmful content, data exposure) from an English-centric "translate-then-detect" design to a single native multilingual model. The article is product and engineering news and reports no incident, vulnerability or threat actor activity. Tenable built an agentic translation pipeline to generate label-preserving training and benchmark data in nine additional languages. It reports roughly 20-800x faster detection per item and about an order of magnitude lower cost than the legacy approach.
Detalhes técnicos
- The pipeline has three stages: a frontier LLM produces a label-aware transcreation, a panel of independent judges scores fluency, meaning integrity and label preservation, and only translations that clear every check are kept.
- Two recovery loops: a refinement loop for failed translations, and a refusal fallback that hands sensitive items to a controlled open-source small language model, which feeds the same quality gate.
- Tenable fine-tuned a compact, already multilingual pretrained model rather than training one from scratch. For some tasks English-only training data generalized to other languages, so translation was needed only for languages where the model underperformed.
- Reported results: detection in tens of milliseconds versus seconds, memory under 1 GB versus 12-14 GB, and a one-time data generation cost under $10,000.
- AI-threat detection recall at a 0.005% false positive rate was 0.93 versus 0.88 for translate-then-detect. Harmful-content recall at a 0.02% false positive rate was 0.65 versus 0.50.
- Supported languages are English, Spanish, German, French, Italian, Portuguese, Mandarin Chinese, Japanese, Arabic and Hindi. Tenable says wider coverage of thousands of other languages is promising but not validated.
- These figures are vendor-reported and not independently verified.
Linha do tempo
-
Tenable details native multilingual AI threat detection in 10 languages for Tenable One AI Exposure
ler a análise desta fonte ocultar
Tenable describes how it moved Tenable One AI Exposure detectors (jailbreaks, prompt injection, harmful content, data exposure) from an English-centric "translate-then-detect" design to a single native multilingual model. The article is product and engineering news and reports no incident, vulnerability or threat actor activity. Tenable built an agentic translation pipeline to generate label-preserving training and benchmark data in nine additional languages. It reports roughly 20-800x faster detection per item and about an order of magnitude lower cost than the legacy approach.
Technical Details
- The pipeline has three stages: a frontier LLM produces a label-aware transcreation, a panel of independent judges scores fluency, meaning integrity and label preservation, and only translations that clear every check are kept.
- Two recovery loops: a refinement loop for failed translations, and a refusal fallback that hands sensitive items to a controlled open-source small language model, which feeds the same quality gate.
- Tenable fine-tuned a compact, already multilingual pretrained model rather than training one from scratch. For some tasks English-only training data generalized to other languages, so translation was needed only for languages where the model underperformed.
- Reported results: detection in tens of milliseconds versus seconds, memory under 1 GB versus 12-14 GB, and a one-time data generation cost under $10,000.
- AI-threat detection recall at a 0.005% false positive rate was 0.93 versus 0.88 for translate-then-detect. Harmful-content recall at a 0.02% false positive rate was 0.65 versus 0.50.
- Supported languages are English, Spanish, German, French, Italian, Portuguese, Mandarin Chinese, Japanese, Arabic and Hindi. Tenable says wider coverage of thousands of other languages is promising but not validated.
- These figures are vendor-reported and not independently verified.