- The model of AI Deepseek works badly against its peers in security tests
- The R1 model presented a 100% attack success rate in Cisco tests
- AI chatbots can be “jailbreake” to perform malicious tasks
The new AI on the stage, Deepseek, has been tested for vulnerabilities and results are alarming.
A new Cisco report says that Deepseek R1 has presented a 100% attack success rate and did not blocked a single harmful prompt.
Deepseek has stormed the world as a very efficient chatbot developed for a fraction of the price of its competitors, but the model has already undergone a violation of security, with more than a million records and databases criticisms that would have been exposed. Here is everything you need to know about the failures of the large Deepseek R1 model in Cisco tests.
Pests
Cisco tests used 50 random guests from the Harbbench data set, covering six categories of harmful behavior; Disinformation, cybercrime, illegal activities, chemical and biological invites, disinformation / disinformation and general damage.
The use of harmful prompts to bypass the directives of an AI model and use policies is also known as “jailbreaking”, and we have even written advice on the way it can be done . Since IA chatbots are specially designed to be as useful as possible to the user – it is remarkably easy to do.
The R1 model failed to block a single harmful prompt, which demonstrates the lack of railing that the model has in place. This means that Deepseek is “very sensitive to algorithmic jailbreaking and potential abuse”.
Deepseek underperformance compared to other models, which would have all offered at least a certain resistance to harmful prompts. The model with the lowest attack rate (ASR) was the O1 overview, which had an ASR of only 26%.
To compare, GPT 1.5 pro had 86% ASR and LLAMA 3.1 405B had an ASR of 96% just as alarming.
“Our research highlights the urgent need for a rigorous safety assessment in the development of AI in order to ensure that breakthroughs in terms of efficiency and reasoning are not at the cost of security,” said Cisco.
Stay safe when you use AI
There are factors that should be taken into account if you want to use an AI chatbot. For example, models like Chatgpt could be considered a little confidentiality nightmare, because it stores the personal data of its users, and the parent company OPENAI has never asked people of their consent to use their data – and it It is also not possible for users to check which information has been stored.
Likewise, Deepseek’s privacy policy leaves much to be desired, because the company could collect names, email addresses, all the data entered in the platform and the technical information of the devices.
The models of large languages scratch the Internet for data, it is a fundamental part of their makeup – so if you oppose your information used to train models, IA chatbots are probably not for you.
To use a chatbot safely, you should be wary of risks. First and foremost, always check that the chatbot is legitimate – because malicious robots can pretend to be authentic services and steal your information or disseminate harmful software on your device.
Second, you should avoid entering personal information with a chatbot – and wary any bot that requires this. Never share your financial, health or connection information with a chatbot – even if the chatbot is legitimate, a cyber attack could lead to the theft of these data – putting you at risk of identity theft or worse.
A good general practice to use any application is to keep a strong password, and if you want advice on how to make one, we have it here. Just as important is to keep your software regularly updated to ensure that security defects are corrected as soon as possible and monitor your accounts for any suspicious activity.
