Cisco Report Claims that DeepSeek is Vulnerable to Attacks

DeepSeek has become the new buzzword in the AI space thanks to its advanced capabilities and free access. However, a recent report from Cisco shows that it has a 100% attack success rate and failed to block a single harmful prompt. This is quite alarming, as the DeepSeek app has become a top-rated free app on Apple’s App Store in the US. 

“Our findings suggest that DeepSeek’s claimed cost-efficient training methods, including reinforcement learning, chain-of-thought self-evaluation, and distillation, may have compromised its safety mechanisms. DeepSeek R1 lacks robust guardrails compared to other frontier models, making it highly susceptible to algorithmic jailbreaking and potential misuse.” Cisco blog post.

Key Findings of the Cisco Report 

The researchers tested DeepSeek R1 using 50 carefully crafted “jailbreak” prompts designed to bypass its safety measures. Shockingly, the model generated harmful responses to every single prompt, which included misinformation, instructions for illegal activities, and guidance on cybercrime. 

In contrast to DeepSeek’s performance, other AI models showed some level of resistance. For instance, OpenAI’s o1 system only responded to 26% of harmful prompts, while Anthropic’s Claude 3.5 Sonnet had a 36% attack success rate. 

What Experts are Saying? 

Jeetu Patel, Cisco’s EVP & Chief Product Officer emphasized the importance of robust networking and security measures as DeepSeek becomes more widely available on platforms like AWS Bedrock. 

Zeus Kerravala from ZK Research expressed skepticism about DeepSeek’s claims of cost-effective performance without compromising security. He stated,” So, is it cheaper to train? It looks like it, but at what cost? Sometimes, the lowest cost is the most expensive.” 

What are the Implications? 

It’s important to keep in mind that this is not the first time DeepSeek has been accused of its AI model. Earlier, there were reports of a critical database leak that exposed over one million records, including system logs, user prompts, and API tokens. The app has also been criticized for promoting Chinese propaganda. 

Due to security and privacy issues, the US Navy has banned the use of DeepSeek on government-issued devices. On a similar note, Italy has also banned the app, citing fears of data privacy concerns.

As of writing this article, DeepSeek has not officially responded to these findings. However, industry experts and analysts await clarification on how the company plans to address these critical security vulnerabilities.