Anthropic CEO says DeepSeek was ‘the worst’ on a critical bioweapons data safety test

Anthropic chief executive officer Dario Amodai is concerned about the Chinese AI Organization Competitive DEPSEC It took the Silicon Valley in the storm With The R1 ModelThe And its anxiety can be more serious than the common people raised about the repayment of DEPSEC user’s data to China.

In An interview Jordan Snyder’s Chinatalk Podcast said Amodi that DIPSEC had created rare information about biopone in the security test conducted by anthropological.

Amodai claimed that the acting of the Dipsic “was basically the worst of any model”. “There was no block of it against the making of this information.”

Amodai said that it was part of the assessment that anthropologists were regularly driven to various AI models to evaluate their potential national security risks. His team sees whether his team can generate information related to bioPones that are not easily available in Google or textbooks. Anthropological positions yourself as the AI ​​Foundational Model Suppliers It takes security seriously.

Amodai said that he did not seem “literally dangerous” in the supply of diploma models today, but they may be in the near future. Although he praised the Diploma team as a “talented engineer”, he advised the agency to “take this AI protection consideration seriously.”

Amodeai is also there Supported strong export control Chips in chips, referring to the concern that they can give China’s military an end.

The DIPSEC model did not clarify in the interview with the ethnographic tested Chinatala, nor did he give more technical details about these tests. Anthropologists instantly did not respond to a request to comment on TechCrunch. Deepsek did not.

The rise of the Dipsic has expressed concern over its protection elsewhere. For example, Cisco Protection Researchers Said last week Failed to block any harmful requests in DEPSEC R1 in its security test, 100% Jailbreak has achieved success rate.

Cisco did not mention the bioewaipon, but said that it was able to get a dipsek to generate harmful information about cybercrime and other illegal activities. Although it is mentioned, Meta’s Lama -1.5-6B and OpenAI GPT -4 OO also had a high failure rate of 96% and 86% respectively.

It is still seen whether the anxiety of this national protection will make a serious tooth within the dipsic Accepted quickly. The company preferred AWS And Microsoft Amazon is sarcastically sufficient, tutted on their cloud platforms to integrate the R1K publicly on the platform EthnicThe

On the other hand, There is a growing list Countries, OrganizationAnd especially the US Navy and the government agencies like it Pentagon That DIPSEC has begun to ban.

Time will tell if these efforts will be captured or the global rise of the diplosh will continue. Whatever the case, Amodai says that he has considered DEPSEC as a new contestant that is at the level of the top AI companies in the United States.

“The new truth here is a new contestant,” he told Chinatal. “AI – anthropologists, OpenAE, Google, probably Meta and Jai – DIPSEC is probably being added to that section in large companies that can provide AI training.”