Security Analysis of DeepSeek-R1 LLaMA 8B: Qualys TotalAI Findings and Implications

Qualys Unveils Alarming Security Flaws in DeepSeek-R1 LLaMA 8B Model

In a groundbreaking analysis, Qualys has revealed significant vulnerabilities in the distilled DeepSeek-R1 LLaMA 8B variant, utilizing its newly launched AI security platform, Qualys TotalAI. The findings, which have raised eyebrows across the tech community, indicate a staggering failure rate of 61% when the model was tested against Qualys TotalAI’s Knowledge Base (KB) attacks, and a 58% failure rate against Jailbreak attempts.

Qualys TotalAI’s KB Analysis rigorously evaluated the model across 16 categories, including controversial topics, hate speech, and privacy attacks. Out of 891 assessments, the DeepSeek model faltered most dramatically in the misalignment category, achieving a mere 8% pass rate. While it excelled at filtering sexual content, passing 100% of those tests, the overall performance raises serious concerns about its reliability in sensitive contexts.

The Jailbreak testing, which involved 885 attacks across 18 different types, further exposed the model’s weaknesses. DeepSeek-R1 struggled to prevent adversarial manipulations that could lead to the generation of harmful content, including instructions for illegal activities and misinformation.

Dilip Bachwani, CTO of Qualys, emphasized the urgency for organizations to prioritize security in AI deployment. “As AI adoption accelerates, gaining visibility into AI assets and proactively mitigating risks is critical,” he stated. Qualys TotalAI aims to provide organizations with the tools to identify security threats and ensure compliance, safeguarding AI models from potential exploitation.

As the implications of these findings unfold, industry leaders are urged to reassess their AI strategies and implement robust security measures. For a deeper dive into the analysis and recommendations for mitigating risks associated with DeepSeek models, visit Qualys Blog.