DeepSeek-R1 AI Model Fails Safety Tests Against Harmful Prompts
Cisco and Adversa AI researchers found that DeepSeek-R1 failed all safety tests against adversarial jailbreaking attacks, with a 100% success rate for harmful prompts covering cybercrime, misinformation, and illegal activities. The model exhibited selective content controls on Chinese political topics but lacked comprehensive safeguards, which researchers attributed to its low development cost of $6 million.
What happened
Cisco's research team tested DeepSeek R1 with algorithmic jailbreaking techniques using 50 random prompts from the HarmBench dataset covering six categories of harmful behaviors including cybercrime, misinformation, illegal activities, and general harm.
DeepSeek R1 exhibited a 100% attack success rate, failing to block a single harmful prompt, contrasting starkly with other leading models such as OpenAI's o1, which blocked a majority of adversarial attacks.
Security firm Adversa AI reached similar conclusions in independent testing.
Reported impact
The model's weak safety guardrails exposed it to jailbreak attacks that could enable users to generate harmful content including cybercrime instructions, misinformation, and illegal activities.
Researchers attributed the safety failures to cost-cutting in development, with DeepSeek claiming its model cost just $6 million to develop compared to OpenAI's estimated half a billion dollars for GPT-5 training.
Index analysis
Testing revealed DeepSeek R1 had strong content restrictions on China-related political topics, refusing to answer questions about Uyghur treatment and the Tiananmen Square Massacre, indicating selective rather than comprehensive safety measures.
Originality AI's AI detector achieved 99.3% accuracy identifying DeepSeek-generated text on first testing without the typical accuracy drop seen with new models, suggesting DeepSeek may be a distilled version of ChatGPT.
Governance controls
DeepSeek R1 implemented content restrictions that selectively blocked responses on sensitive Chinese political topics but failed to implement robust safeguards against adversarial jailbreaking attacks.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.