DeepSeek’s R1 AI is 11 times more likely to be exploited by cybercriminals than other AI models – whether that’s by producing harmful content or being vulnerable to manipulation.
This is a worrying finding from new research conducted by Enkrypt AI, an AI security and compliance platform. This security warning adds to the ongoing concerns following last week’s data breach that exposed over one million records.
China-developed DeepSeek sent shockwaves throughout the AI world since its January 20 release. About 12 million curious users worldwide downloaded the new AI chatbot in the space of two days, marking a growth even faster than ChatGPT. Widespread privacy and security concerns have, however, prompted quite a few countries to either begin investigating or banning, in some capacity, the new tool.
Harmful content, malware and manipulation
The team at Enkrypt AI performed a series of tests to evaluate DeepSeek’s security vulnerabilities, such as malware, data breaches, and injection attacks, as well as its ethical risks.
The investigation found the ChatGPT rival “to be highly biased and susceptible to generating insecure code,” experts noted, and that DeepSeek’s model is vulnerable to third-party manipulation, allowing criminals to use it for developing chemical, biological, and cybersecurity weapons.
Nearly half of the tests conducted (45%) bypassed safety protocols in place, generating criminal planning guides, illegal weapons information, and terrorist propaganda.
Worse still, 78% of the cybersecurity checks successfully tricked DeepSeek-R1 into generating insecure or malicious codes. These included malware, trojans, and other exploits. Overall, experts found the model to be 4.5 times more likely than its Open-AI counterpart to be manipulated by cybercriminals to create dangerous hacking tools.
“Our research findings reveal major security and safety gaps that cannot be ignored,” said Sahil Agarwal, CEO of Enkrypt AI, commenting on the findings. “Robust safeguards – including guardrails and continuous monitoring – are essential to prevent harmful misuse.”
🚨 Are Distilled DeepSeek Models Less Safe? Early Signs Point to Yes. 🚨Our latest findings confirm a concerning trend: distilled AI models are more vulnerable—easier to jailbreak, exploit, and manipulate. 📄 Read the Paper: https://t.co/nzdcR82J8M 🔍 Key Takeaways 🔹… pic.twitter.com/IFCjLYxBWBJanuary 30, 2025
As mentioned earlier, at the time of writing DeepSeek is under scrutiny in many countries worldwide.
While Italy was the first to launch an investigation into its privacy and security last week, many EU members have followed suit so far. These include France, the Netherlands, Luxembourg, Germany, and Portugal.
Some of China’s neighboring countries are getting worried, too. Taiwan, for example, has banned all government agencies from using DeepSeek AI. Meanwhile, South Korea initiated a probe into the service provider’s data practices.
Unsurprisingly, the US is also taking aim at its new AI competitor. As NASA blocked DeepSeek usage on federal devices – CNBC reported on Friday, January 31, 2025 – a proposed law could now outright ban the use of DeepSeek for all Americans who could risk million-dollar fines and even prison time for using the platform in the country.
All in all, Agarwal from Encrypt AI said: “As the AI arms race between the US and China intensifies, both nations are pushing the boundaries of next-generation AI for military, economic, and technological supremacy.
“However, our findings reveal that DeepSeek-R1’s security vulnerabilities could be turned into a dangerous tool – one that cybercriminals, disinformation networks, and even those with biochemical warfare ambitions could exploit. These risks demand immediate attention.”
https://cdn.mos.cms.futurecdn.net/RFHhebaafuggVviVNzSsy-1200-80.jpg
Source link
chiara.castro@futurenet.com (Chiara Castro)
