DeepSeek’s latest AI model, R1, has raised concerns over its vulnerability to jailbreaking, making it easier to manipulate for harmful purposes. According to The Wall Street Journal, tests revealed that the model could generate dangerous content, including plans for a bioweapon attack and strategies for promoting self-harm among teenagers.
Sam Rubin, senior vice president at Palo Alto Networks’ Unit 42, confirmed that R1 is more susceptible to such exploits compared to other AI models. The Journal conducted its own tests and successfully prompted R1 to create a social media campaign designed to manipulate teens by exploiting their emotions.
The chatbot was also tricked into providing instructions for a bioweapon attack, drafting a manifesto supporting Hitler, and writing a phishing email embedded with malware. When the same prompts were given to ChatGPT, it refused to comply, highlighting the security gap in DeepSeek’s model.
This discovery adds to existing concerns about DeepSeek’s content restrictions, as it reportedly avoids discussions on sensitive topics like Tiananmen Square and Taiwanese autonomy. Additionally, Anthropic CEO Dario Amodei recently stated that DeepSeek performed the worst in a safety test focused on bioweapons.
The findings raise questions about the company’s approach to AI safety, particularly as it gains traction in Silicon Valley and global markets. While AI technology continues to evolve, ensuring robust security measures is crucial to prevent misuse. DeepSeek’s apparent shortcomings in this area may prompt further scrutiny from regulators and AI researchers.
