New AI Trust score reveals DeepSeek leads in sensitive information disclosure

Recent evaluations reveal that Chinese AI models, particularly DeepSeek, are outperforming their American counterparts, such as Meta's Llama, in critical areas including the disclosure of sensitive information.

This finding comes from the newly introduced AI Trust Score developed by Tumeryk, which assesses AI models across nine key factors, including sensitive information disclosure, insecure output handling, security, toxicity, etc. Tumeryk's AI Trust Manager is designed to assist security professionals to better ensure that their AI deployments are secure and compliant, while also identifying vulnerabilities within their AI applications and monitor the performance, in real-time. It can also recommend actionable steps to enhance the system's security and compliance.

DeepSeek's model, known as DeepSeek NIM, has impressively garnered a score of 910 in the sensitive information disclosure category, surpassing Anthropic Claude's score of 687 and Meta Llama's score of 557. These results indicate that the landscape of AI development is evolving, challenging existing perceptions about foreign models and their safety and compliance standards.

Betanews reports that according to the evaluations, DeepSeek and other Chinese models are reportedly exhibiting higher safety and compliance standards than previously understood. Moreover, these models run on US-based platforms like NVIDIA and SambaNova, ensuring data security while abiding by international regulations. This confluence of performance and compliance is pivotal for companies looking to adopt AI technologies securely and ethically.

As the AI industry continues to evolve, such objective, data-driven evaluations will be critical in shaping trust and transparency among users and developers alike.