DeepKeep Outperforms Meta and Nvidia in Multilingual AI Security Tests

The research highlights a critical vulnerability in modern guardrails: a heavy reliance on English. A study from Brown University previously demonstrated that GPT-4 could be coerced into fulfilling harmful requests 79% of the time simply by translating prompts into low-resource languages. While many current security tools attempt to bridge this gap by translating inputs into English, this process often introduces latency and strips away essential context, creating significant blind spots for multinational organizations.

DeepKeep bypasses this bottleneck through a cognition-based analysis that evaluates the semantic intent of prompts directly in their original language. During benchmark testing using datasets like SafeGuard and Wild Jailbreak, the platform achieved F1 scores nearing 0.98 for prompt injection detection. Notably, the system maintains this performance using a 400-million parameter architecture. This efficiency allows for lower latency compared to multi-billion parameter models that currently dominate the market.

"AI security has largely been built around the assumption that prompts are written in English," said Yossi Altevet, CTO and Co-Founder at DeepKeep. "Security models must understand intent across languages, not just words." By handling mixed-language queries seamlessly and identifying PII without needing translation, the platform addresses the specific security risks inherent in global AI deployment.