“Gelbrex persists only because it is almost impossible to eliminate them completely – buffer overflow weaknesses in software (which have been present for more than 40 years) or SQL Injection Doshas in web applications (which protection for more than two decades Like teams), like “Alex”, “Alex,” Alex, “Alex,” Alex, “Alex,” Alex, “Alex,” Alex, “Security Firm APversa AI's CEO Polycov told.
Cisco's Sampat argues that as companies use more types of AIs in their applications, risks are increased. “It starts to become a big thing when you start putting these models into important complex systems and those gelbreaks suddenly result in downstream things that increase liability, increase business risk, all kinds of issues for enterprises for enterprises Increases, “Sampat says.
The researchers in Cisco made their 50 randomly selected indications from a famous library of standardized evaluation to test the R1 of Deepsek, known as the hormorben. He tested signs from six harmbench categories including general disadvantages, cybercrime, misinformation and illegal activities. He checked the local running model on machines instead of Deepsek's website or app, which Send data to china,
In addition, researchers say that they have also seen some possible results from R1's testing, including more involved, non-linguistic attacks with things like Cyrillic characters and analog script to achieve code execution. Efforts have been made. But for his early tests, Sampat says, his team wanted to focus on the conclusions that were commonly stems from a recognized benchmark.
The performance of other models in Cisco also included a comparison of the performance of R1 against hormabench signals. And some, like Meta Lama 3.1Deepsek's R1 was almost severely faltering. But Sampat has been emphasized that R1 of Deepsek is a distinctive Argument modelWhich takes more time in generating answers, but draws more complex processes to try to give better results. Therefore, Sampat argues, with the best comparison OPENAI's O1 logic modelWhich performed best in all the tested models. (Meta did not immediately respond to the request of the comment).
Polykov of APversa AI, explains that Deepsek appears to detect and reject some famous jailbreak attacks, saying “it seems that these reactions are often copied from Openai's dataset.” However, Polycov says that in four different types of gelbrex tests of his company-from spectacular people to code-based tricks-deepsecch sanctions can be easily bypassed.
Polycova says, “Each method worked innocently.” “It is even more worrying that these novels are not 'zero-day' not gelbreaks-many have been known for years,” he said, he claimed that the model said that the model said more with some instructions around psychidelix. Was seen going in depth. Other models make.
Polycov says, “Deepsek is just another example of how every model can be broken – it's just how much you try. Some attacks can be patched, but the surface of the attack is endless , “Polycov is called. “If you are not constantly making your AI red, then you are already compromising.”