DeepSeek AI model fails security tests and spreads misinformation
Bulgarian and international researchers tested DeepSeek AI models and found serious vulnerabilities in cybersecurity and bias management, including susceptibility to jailbreaking attacks that could cause life-threatening harm. The models were modified from open-source versions by Meta and Alibaba, and testing revealed they rank lowest among competitors in preventing prompt injection and prompt leakage attacks.
What happened
Researchers at INSAIT (Sofia University) and LatticeFlow AI tested DeepSeek R1 8B and R1 14B models and found serious vulnerabilities in cybersecurity and content management. The models ranked lowest among tested competitors including OpenAI, Meta, Google, Anthropic, and Mistral AI in cybersecurity performance.
DeepSeek models were particularly vulnerable to goal hijacking and prompt leakage attacks, according to LatticeFlow. Security researchers found the models could be manipulated through jailbreaking techniques to provide step-by-step instructions for creating explosives or setting fires, actions the models should have prevented.
The models showed significantly higher bias compared to their base versions from Meta's Llama 3.1 and Alibaba's Qwen 2.5, generating imbalanced responses on sensitive topics and spreading misinformation across various subjects. DeepSeek modified the open-source models during development, and the modifications appear to have compromised key safety features present in the original versions.
Reported impact
The vulnerabilities undermine the reliability of the models and make their use in business environments risky, according to the research. The models' susceptibility to jailbreaking could allow users to bypass safety guardrails and obtain harmful instructions.
CEO Alex Polakov of Adversa AI stated that 'DeepSeek is completely undefended against all jailbreaking approaches, while reasoning models from OpenAI and Anthropic have become much more secure compared to their older non-reasoning versions tested last year.'
Index analysis
The models rank below average on bias assessment, which measures how objective, neutral and fair responses are regarding different social, cultural, ethnic, gender and political groups. DeepSeek models performed significantly worse than their base versions in this evaluation.
Governance controls
The research notes that DeepSeek models have serious gaps in compliance with the EU AI Act. DeepSeek performed well in limiting toxic content, exceeding even their base models, but lagged in cybersecurity and bias management.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.