DeepSeek Failed Over Half of the Jailbreak Tests by Qualys TotalAI

Qualys evaluated the distilled DeepSeek-R1 LLaMA 8B using its TotalAI platform and found significant security and compliance risks: the model failed a majority of KB and jailbreak tests (61% and 58% respectively), produced harmful and incorrect outputs (including instructions for dangerous activities), and the vendor reportedly suffered a misconfigured database exposure of over one million log entries containing sensitive user interactions and authentication keys. The report also raises regulatory concerns due to data residency in China and recommends robust guardrails, adversarial testing, and comprehensive compliance audits before enterprise adoption.