#Israel #Tel Aviv #AI Security #Multilingual Performance #DeepKeep

DeepKeep Elevates AI Security with Multilingual Capabilities

In today's interconnected world, the need for robust AI security solutions has never been more critical, especially as organizations embrace AI technologies across diverse languages. DeepKeep, a leading end-to-end AI security platform, has demonstrated significant strides in this arena through a recent benchmark study. The findings indicate a remarkable enhancement in DeepKeep's multilingual AI security performance, setting a new standard in the industry.

The research results underline that many existing AI security systems have primarily been designed to handle English-language prompts, creating vulnerabilities in a global context. A notable study from Brown University revealed that AI systems like GPT-4 can interact with harmful requests in low-resource languages to an alarming extent—79% of the time—compared to less than 1% with English inputs. This stark contrast highlights an enticing opportunity for malicious actors to exploit weaknesses in AI security by issuing prompts in multiple languages, thereby bypassing traditional guardrails.

DeepKeep's innovative approach tackles this issue head-on by deploying cognition-based analysis, which directly evaluates the semantic meaning of prompts and responses irrespective of the language in which they were issued. This method eliminates the need for initial translation into English, thus ensuring that the essence of the communication is preserved while simultaneously reducing latency and enhancing response times.

One of the standout features of DeepKeep's security model is its ability to seamlessly manage mixed-language prompts—a scenario increasingly prevalent in corporate environments where users often intersperse various languages within single AI queries. This adaptability is critical as businesses seek to implement AI at a global scale.

In comparative benchmark tests involving widely used open-source models, including Meta’s LLaMa Prompt Guard and Nvidia’s NeMo, DeepKeep emerged as the leader in detecting prompt injection attempts and safeguarding personally identifiable information (PII). The evaluation utilized several respected prompt injection datasets, including SafeGuard and Wild Jailbreak, with test sets translated into a wide array of languages such as Japanese, German, Spanish, and Portuguese.

Consequently, DeepKeep achieved exceptional F1 scores nearing 0.98 in prompt injection detection, all while significantly minimizing false negatives compared to translation-dependent models. This high performance is indicative of DeepKeep's commitment to maintaining safety without compromising efficiency, even within the intricacies of multilingual environments.

As Yossi Altevet, CTO and Co-Founder, emphasized, "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." This statement underscores the necessity for a paradigm shift in the security models employed for AI interactions, a shift that DeepKeep has adeptly navigated.

Furthermore, DeepKeep's multilingual classifier operates efficiently with a model containing approximately 400 million parameters. This design contrasts sharply with other models that typically rely on multi-billion-parameter architectures, rendering DeepKeep’s solution faster and more responsive—qualities vital for enterprise deployments.

DeepKeep's multilingual security solution forms a core part of the company's extensive array of enterprise AI security offerings. These solutions empower organizations to leverage AI technologies confidently, without sacrificing safety, control, or trust.

In summary, as enterprises increasingly deploy AI tools across global platforms, understanding language intent rather than mere words becomes crucial. With its strategically developed multilingual capabilities, DeepKeep is poised to redefine standards in AI security, ensuring that users can trust their interactions with these transformative technologies. For more insights on DeepKeep's groundbreaking multilingual AI security research, visit DeepKeep's official website.