
DeepSeek's Guardrails Blocked 0 of 50 Bad Prompts, Which Is Bold for a Chatbot
Ourtake
Zero blocked, one hundred percent attack success. The safety budget apparently went to the cheap part.
What happened
Security researchers from Cisco and the University of Pennsylvania tested DeepSeek's R1 reasoning model with 50 malicious prompts and reported it did not detect or block any of them. Adversa AI separately reported that DeepSeek's restrictions could be bypassed with jailbreak methods. DeepSeek did not respond to a request for comment.
Sources
- Wired · DeepSeek’s Safety Guardrails Failed Every Test Researchers Threw at Its AI Chatbot (opens in a new tab)
- blogs.cisco.com · Evaluating Security Risk in DeepSeek and Other Frontier Reasoning Models (opens in a new tab)
- Futurism · DeepSeek Failed Every Single Security Test, Researchers Found (opens in a new tab)