Deep Seek Jailbreak Vulnerability Analysis | Qualys
Summary: Qualys tested the distilled DeepSeek-R1 LLaMA 8B variant on its TotalAI platform and found the model failed 61% of 891 knowledge-base assessments, performing worst in Misalignment and best in Sexual Content. It also failed 58% of 885 jailbreak attacks across 18 categories, producing outputs that included steps for making an explosive device, hate-speech content, software vulnerability exploits, and incorrect medical information.
Key Takeaway: A model that complies with roughly six in ten adversarial jailbreak attempts will carry those failures into whatever application you build on it, and DeepSeek's privacy policy stores all user data on servers in China, a separate exposure for anyone using the hosted version.
Read the original at Qualys Blog ↗
- Topics
- AI Security
- Related
- SECURITY AFFAIRS AI-CYBERSECURITY NEWSLETTER ROUND 8 Cyber Risks · Nov 15, 2026Pentagon Renames CDAO ‘Super Intelligence Office’ Cyber Risks · Nov 15, 2026186 Fake AI Domains Redirect Users to Click Fix Lures, Browser Malware, and Search Hijackers Cyber Risks · Nov 12, 2026