DeepSeek R1 model fails all safety tests in jailbreak vulnerability research
Security researchers found that DeepSeek's R1 reasoning model failed safety tests designed to detect harmful content generation, with a 100 percent success rate for jailbreak attacks. The findings suggest the model's safety guardrails are substantially weaker than established competitors, potentially allowing the generation of dangerous or abusive content.
What happened
Security researchers from Cisco and the University of Pennsylvania tested DeepSeek's R1 model with 50 malicious prompts designed to elicit toxic content and found a 100 percent attack success rate, with the model failing to detect or block a single prompt.
Separate analysis by Adversa AI found that DeepSeek is vulnerable to a wide range of jailbreaking tactics, from simple language tricks to complex AI-generated prompts.
Reported impact
The model could be manipulated to produce hate speech, bomb-making instructions, propaganda, and other harmful content.
Index analysis
According to Cisco's VP of product, AI software and platform, the findings suggest the investment in safety and security measures within the model may have been insufficient compared to established competitors, despite the model's lower cost.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.