Anthropic CEO says DeepSeek was ‘the worst’ on a critical bioweapons data safety test | TechCrunch
The general director of Anpropic Dario Amodea is worried about the competitor of Deepseek, the Chinese art company that stormed a silicone valley with its R1 model. And his fears can be more serious than typical, grown about Deepseek, sending user data back to China.
In an interview with the Jordan Schneider’s Chinatalk Amodei podcast, DeepSeek generated rare information about bio -Apone in testing safety controlled by ANTROPIC.
The performance of DeepSeek was “the worst of mostly any model that we once tested,” said Amodee. “He had absolutely no blocks against the creation of this information.”
Amodei said that this was part of the assessment that the anthropic regularly works on various AI models to evaluate their potential risks of national security. His team is watching whether the models can generate information related to bioapons, which is not easy to find in Google or in textbooks. Anthropic positions itself as a supplier of the fundamental models of AI, which is seriously security.
Amodea said that today he did not think that Deepseek models are “literally dangerous” in providing rare and dangerous information, but they could be in the near future. Although he praised the DeepSeek team as “talented engineers”, he advised the company “seriously treat these artificial intelligence safety reasons.”
Amodei also supports strong export controls on chips in China, citing fears that they can give military China an advantage.
Amodea did not specify in an interview with the Chinese Chinese point that Deepseek Anpropic tested, and he did not give any more technical details about these tests. Anthropic did not immediately respond to a request about the comments from TechCrunch. Not Deepseek.
Deepseek growth caused concern about her safety in other places. For example, last week, Cisco security researchers said that Deepseek R1 could not block any harmful tips in their safety tests, having achieved 100% success in Jailbrack.
Cisco did not mention Bioweapons, but said that he could get Deepseek to get harmful information about cybercrime and other illegal actions. It is worth noting, however, that META’s LLAMA-3,1-405B and GPT-4O Openai also had a high level of failures 96% and 86%, respectively.
It is still unknown whether such problems of the safety of a serious dent in the rapid adoption of Deepseek will be. Companies such as AWS and Microsoft publicly advertised the integration of R1 into their cloud platforms – by the irony of fate, given that Amazon is the largest investor Anpropic.
On the other hand, there is a growing list of countries, companies and especially government organizations, such as the US Navy and the Pentagon, which began to prohibit Deepseek.
Time will tell if these efforts will take or the global growth of Deepseek will continue. In any case, the Amodea says that he considers Deepseek a new competitor, which is at the level of the best companies in the USA.
“The new fact here is that there is a new competitor,” he said to Chinatalk. “In large companies that can be trained – anthrop, open, Google, perhaps Meta and Xai – now, perhaps, Deepseek is added to this category.”