DeepSeek R1 Vulnerabilities Expose Cybercrime Risks and IP Misuse
Security researchers found that DeepSeek R1 is highly vulnerable to jailbreak attacks and failed all safety tests, while OpenAI allegedly discovered evidence the model was trained on OpenAI's proprietary data. Additionally, a publicly accessible database belonging to DeepSeek exposed internal information including secret keys and conversation histories.
What happened
Cisco security researchers tested DeepSeek R1 against 50 harmful prompts from the HarmBench dataset covering cybercrimality, disinformation, illegal activities and general harms, and found a 100% success rate for jailbreaking attacks—the model failed to block any harmful prompt.
OpenAI stated it had found evidence that DeepSeek used OpenAI's proprietary models to train its own open-source model. Microsoft integrated DeepSeek R1 onto Azure AI Foundry and GitHub despite these accusations.
Wiz Research identified a publicly accessible ClickHouse database belonging to DeepSeek containing over one million rows of data flow history, including conversation history, secret keys, backend details and internal information.
Reported impact
Cisco researchers highlighted that DeepSeek R1's vulnerability contrasts sharply with other leading models such as o1, which blocked a majority of adversarial attacks. Cisco noted DeepSeek's low development budget of $6 million compared to competitors' higher spending may have resulted in reduced safety and security measures.
Index analysis
The jailbreak vulnerability allows the model to provide harmful information on cybercrimality and illegal activities without safeguards. The data exposure disclosed internal operations details and secret credentials accessible to the public.
Governance controls
Microsoft stated that DeepSeek R1 underwent rigorous red team evaluations and security testing, including automated model behavior evaluations and thorough security reviews to mitigate potential risks.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.