Alex said in a Wired email in Web programs (which have plagued security teams for more than two decades).
Cisco’s Sampat argues that the more companies use artificial intelligence in their applications, the risks are reinforced. “When you start putting these models in complex systems, it becomes a big job, and these prisoners suddenly lead to downstream items that increase responsibility, Increases the risk of business, increasing different types for companies. “
Cisco researchers have randomized their 50 selected requests to test R1 Deepseek from a well -known library of standardized evaluations known as Harmbench. They tested six categories of Harmbench, including public injury, cybercrime, misconduct and illegal activities. They worked locally on the devices instead of a website or Deepseek program, which sends data to China.
Beyond that, the researchers say they have also observed the potential results of R1 experimentation with more and non -linguistic attacks using things like Cyrillic characters and scripts appropriate to try to achieve the implementation of the code. But for its initial tests, his team wanted to focus on the findings that are generally known as a criterion, Sampat says.
Cisco also included comparing R1 performance against Harmbench with the performance of other models. And some, like the Llama 3.1 meta, plunged almost as R1 Deepseek. But Sampat emphasizes that the R1 Deepseek is a particular reasoning model, which takes more to produce responses but deals with more sophisticated processes to try to get better results. Therefore, Sampat argues, the best comparison with the O1 Openai reasoning model is the best tested models. (Meta did not respond immediately to the comment’s request).
Polyakov, on the other hand, explains that Deepseek seems to have identified and rejected some famous prison escape attacks and says, “It seems that these answers often only copy from the Openai data collection Become. ” However, Polykov says it can be easily removed from four types of prison escapes-from language-based to DeepSEek limitations.
“Every faulty method worked,” says Poly Icof. “What is even more worrying is that these prisoners are not” zero “-they have been publicly known for years,” he says. It creates another model.
“Deepseek is just another example of how each model is broken. It is just how much you try. Some attacks may be packed, but the level of attack is infinite.” “If you are not constantly reducing your AI, you are already at risk.”
