DeepKeep, the end-to-end AI security platform, today unveiled the results of a new benchmark study showcasing significant improvements in the performance and efficiency of its multilingual AI security solution. The research study compared DeepKeep’s approach to other guardrails and LLM-as-a-Judge methods and demonstrated that DeepKeep’s multilingual capability provides superior accuracy and consistency.
As organizations deploy AI tools across global teams, the prompts and interactions these systems process increasingly span multiple languages. However, many AI security systems today are designed primarily for English-language prompts, and the consequences are measurable. A Brown University study found that translating unsafe inputs into low-resource languages got GPT-4 to engage with harmful requests 79% of the time, versus under 1% in English. As enterprises deploy AI globally, attackers can exploit this gap by issuing malicious prompts in other languages to bypass guardrails. While translation-based security solutions exist, these often introduce latency, lose context, or produce inconsistent results across languages, creating a growing security blind spot for multinational enterprises.
DeepKeep directly analyzes the semantic meaning of prompts and responses across languages using a cognition-based analysis, allowing the system to classify data without translating it into English first. The cognition-based solution can deliver an interpretable response, allowing the system to learn and improve in time. Due to their design, DeepKeep’s guardrails can efficiently handle zero-day attacks. The guardrails are also designed to handle mixed-language prompts seamlessly – a scenario increasingly common in enterprise environments where users combine multiple languages within a single AI query.
In benchmark testing against widely used open source models, including Meta’s LLaMa Prompt Guard and Nvidia’s NeMo, DeepKeep significantly outperformed in detecting prompt injection attempts and Personal Identifiable Information (PII).
The evaluation used several widely recognized prompt injection datasets, including SafeGuard, Wild Jailbreak, and Alpaca, with test sets translated into 12 additional languages including Japanese, German, Spanish, French, Italian, Korean, Dutch, and Portuguese. Across these benchmarks, DeepKeep achieved F1 scores approaching 0.98 in prompt injection detection while significantly reducing false negatives compared to translation-based guardrail models.
The platform maintains consistent security decisions across languages while operating with smaller, more efficient model architecture. DeepKeep’s multilingual classifier operates with a model of roughly 400 million parameters, which is significantly smaller than many guardrail models that rely on multi-billion-parameter architectures, enabling faster inference and lower latency in enterprise deployments.
“AI security has largely been built around the assumption that prompts are written in English or that translation is sufficient. That assumption no longer reflects how enterprises actually use AI,” said Yossi Altevet, CTO and Co-Founder at DeepKeep. “As AI systems are deployed across global teams and markets, security models must understand intent across languages, not just words. That shift requires a fundamentally different approach to how AI interactions are analyzed and protected, and we’re proud to have created a solution that meets this challenge with flying colors.”
DeepKeep’s multilingual security solution is part of the company’s established suite of enterprise AI security solutions that give businesses the confidence to leverage AI without sacrificing safety, control, or trust.
Read Also: LocalImpact Report: 88% of Consumers Have Avoided a Business Due to Negative Reviews
