Tenable Deploys Native Multilingual AI Detectors to Replace Slow Translation
Tenable’s new agentic translation pipeline enables native threat detection in ten languages, removing the latency and cost of prior translation methods.
Key points
- Tenable replaced its English-centric detection with a single multilingual model covering ten languages.
- The new system is 20 to 800 times faster and significantly cheaper than previous translation-based methods.
- An agentic pipeline generates high-fidelity training data while preserving security labels across languages.
Tenable has updated its AI security detectors to process threats natively in ten languages, abandoning its previous reliance on English-only analysis. According to research published by the vendor, this shift addresses the limitations of tools that only detect threats in English, which leave organizations exposed to attacks written in other tongues.
The company built an agentic translation pipeline to generate high-fidelity training and benchmark data across nine additional languages. This system includes a multi-judge quality gate, a refinement loop, and a fallback for model refusal. The resulting single multilingual model detects threats natively in all ten supported languages without requiring a separate translation model during production.
Technical Improvements Over Previous Methods
Tenable’s previous approach used a “translate-then-detect” process that converted inputs to English before analysis. According to Tenable, this method suffered from high latency, with large autoregressive models taking seconds for long inputs. It also incurred recurring costs for every non-English request and required significant operational overhead to maintain the translation infrastructure.
The new native detection model eliminates these downsides. Tenable states the new system is roughly 20 to 800 times faster per item than the previous translation process. It is also less expensive by an order of magnitude, bearing a one-time cost that does not increase with customer usage traffic. This removes the need to serve and scale a heavyweight translation model alongside the detector.
Challenges in Multilingual Data Preparation
Creating reliable multilingual data proved difficult because standard translation methods often distort security context. Tenable found that human expert translation does not scale, while machine translation can strip out malicious intent or create awkward phrasing that degrades training signals. For example, native translation might soften a toxin synthesis request into a benign chemistry query.
Off-the-shelf large language models also presented issues, as they may refuse to translate material they flag as inappropriate, such as prompts about weapons or self-harm. This refusal causes likely threat samples to be excluded from the dataset. Tenable’s agentic pipeline addresses these issues by ensuring labels survive the translation process without flipping or defanging the security-relevant language.
What should you do?
Security professionals should verify that their AI security tools support native detection in the languages their users employ. According to Tenable, prompt attacks in Spanish, Arabic, or Hindi are as dangerous as those in English. Organizations relying on English-only detectors remain at risk from non-English attacks. Users should monitor for updates to their security stacks that introduce native multilingual capabilities to reduce latency and detection errors.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: Tenable
- Audit your current AI security tools to confirm they detect threats natively in all languages used by your organization, rather than relying on translation.
- Monitor vendor announcements for updates that replace translation-based detection with native multilingual models to reduce latency and cost.
- Review your detection logs for potential gaps where non-English inputs may have been misclassified due to translation errors or refusal.
- Ensure your security strategy accounts for the fact that prompt attacks in any language pose the same level of risk as English-based attacks.
Step-by-step guide: Social Engineering Warning Signs and How to Respond
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
Why did Tenable move away from translating inputs to English?
Translation added significant latency, cost, and operational overhead, while also risking the loss of security context during the conversion process.
How many languages does the new model support?
The new multilingual model detects threats natively in ten supported languages, including English and nine additional languages.
What is an agentic translation pipeline?
It is a system that generates high-fidelity, label-preserving training data across multiple languages using quality gates and refinement loops.
