Tenable Updates AI Detectors to Scan Threats in Ten Languages Natively

Customers using Tenable One AI Exposure now have protection against threats written in ten languages, including Spanish, Arabic, and Hindi. Previously, the tool relied on translating non-English inputs into English before scanning them for policy risks, harmful content, or AI threats. That approach left organizations vulnerable to attacks crafted in other languages, according to Tenable.

The company replaced that workflow with a single multilingual model that detects threats natively. This change eliminates the latency and cost associated with running large translation models before every detection. Tenable states the new system is roughly 20 to 800 times faster per item and reduces expenses by an order of magnitude.

How the detection model changed

Tenable’s previous method involved translating input to English using a large model, which added hundreds of milliseconds of delay for short inputs and seconds for longer ones. This process also risked losing signal, as naive translation could blur meaning or soften malicious intent, degrading detection accuracy. Operational overhead increased because teams had to maintain and scale an additional heavyweight model alongside the detector.

To build the new capability, Tenable created an agentic translation pipeline to generate high-fidelity training data across nine additional languages. This pipeline includes a multi-judge quality gate, a refinement loop, and a fallback for model refusals. The goal was to preserve labels during translation, ensuring that malicious samples did not become benign or borderline in the training set.

Who is affected by this update

Any organization using Tenable One AI Exposure for monitoring AI threats or policy risks is affected by this change. Users who previously encountered delays or missed detections when processing non-English content will see improved performance. The update specifically targets the risk that prompt attacks written in languages like Spanish or Arabic are just as dangerous as those in English, a gap the English-centric detectors previously left open.

Our take on the response

Shifting from a translate-then-detect workflow to native multilingual detection removes a significant bottleneck for global enterprises. That suggests organizations with diverse user bases or international AI deployments will experience fewer false negatives and lower operational costs. The one-time cost structure, which does not grow with usage traffic, also indicates a more scalable approach to handling increasing volumes of non-English data without proportional increases in compute expenses.

For customers and users

Teams should verify that their current Tenable One AI Exposure configurations are updated to leverage the new multilingual model. Check whether your monitoring covers all relevant languages used by your user base or target audience. Review recent alerts for any non-English content that may have been missed or delayed under the previous translation workflow. Be aware that attackers may still attempt to use mixed-language prompts to evade detection, so maintain regular review of your security policies and

What to do and how to stay safe

• Written plans prevent panic-driven errors that worsen security breaches.
• Testing reveals gaps in communication channels before a real crisis occurs.
• Clear roles ensure technical staff act without waiting for executive approval.

Incident response plans turn panic into procedure, limiting damage when defenses fail. Test your plan regularly to ensure it works when you need it most.

Step-by-step guide: Incident Response Plans: How to Stop Damage When Attacks Happen

General security guidance from the DataSafe Radar newsroom. It is not confirmed advice from the organisations named in this story.