Researchers claim DeepSeek easier to jailbreak than rivals

Author: Swagath Bandhakavi | Source: Tech Monitor | Read the full article | Published February 10, 2025

Recent research has revealed that DeepSeek's R1 AI model is notably easier to manipulate than similar models from major companies like OpenAI and Google. Security experts tested DeepSeek's R1 and found it vulnerable to "jailbreaking," a technique used to bypass safety features in AI systems. This means that users could trick the model into providing harmful or restricted information, raising serious concerns about its security.

The tests conducted by various AI security firms showed alarming results. For instance, researchers were able to get the model to generate dangerous content, including instructions for creating weapons and evading law enforcement. While DeepSeek's R1 did have some basic safety measures, these were easily circumvented, allowing the model to produce inappropriate and harmful outputs that other AI systems would refuse to generate.

Despite these vulnerabilities, DeepSeek's AI is being adopted by several major Chinese companies, including automobile and telecom firms. This growing interest in DeepSeek's technology highlights the ongoing challenges in ensuring AI safety, especially as the model is open source, allowing anyone to modify its code. As the debate over AI security continues, the findings about DeepSeek's R1 serve as a critical reminder of the potential risks associated with advanced AI systems.

[Read More]