Anthropic CEO Slams DeepSeek for Failing Bioweapons Data Safety Test

# DeepSeek

Anthropic's CEO, Dario Amodei, is seriously concerned about DeepSeek, the Chinese AI company that's been making waves in Silicon Valley with its R1 model. And his worries go beyond the usual fears about DeepSeek sending user data back to China. In a recent chat on Jordan Schneider's ChinaTalk podcast, Amodei revealed that DeepSeek's AI model generated some pretty scary info about bioweapons during a safety test run by Anthropic. He claimed that DeepSeek's performance was "the worst of basically any model we'd ever tested," saying it had "absolutely no blocks whatsoever against generating this information." Amodei explained that these tests are part of Anthropic's regular evaluations to check if AI models pose any national security risks. They specifically look at whether the models can spit out bioweapons-related info that you can't easily find on Google or in textbooks. Anthropic prides itself on being the AI company that takes safety seriously. While Amodei doesn't think DeepSeek's current models are "literally dangerous" when it comes to providing rare and risky info, he believes they could be in the near future. He praised DeepSeek's team as "talented engineers" but urged the company to "take seriously these AI safety considerations." Amodei has also been a vocal supporter of strong export controls on chips to China, worried that they could give China's military an advantage. In the ChinaTalk interview, Amodei didn't specify which DeepSeek model Anthropic tested or provide more technical details about the tests. Neither Anthropic nor DeepSeek responded right away to TechCrunch's request for comment. DeepSeek's rapid rise has raised safety concerns in other places, too. Last week, Cisco security researchers said that DeepSeek R1 failed to block any harmful prompts in its safety tests, with a 100% jailbreak success rate. While Cisco didn't mention bioweapons, they were able to get DeepSeek to generate harmful info about cybercrime and other illegal activities. It's worth noting, though, that Meta's Llama-3.1-405B and OpenAI's GPT-4o also had high failure rates of 96% and 86%, respectively. It's still up in the air whether these safety concerns will put a serious dent in DeepSeek's rapid adoption. Companies like AWS and Microsoft have been publicly praising the integration of R1 into their cloud platforms – which is kind of ironic, considering Amazon is Anthropic's biggest investor. On the flip side, a growing number of countries, companies, and especially government organizations like the U.S. Navy and the Pentagon have started banning DeepSeek. Only time will tell if these efforts gain traction or if DeepSeek's global rise will keep going strong. Either way, Amodei considers DeepSeek a new competitor that's on par with the U.S.'s top AI companies. "The new fact here is that there's a new competitor," he said on ChinaTalk. "In the big companies that can train AI – Anthropic, OpenAI, Google, perhaps Meta and xAI – now DeepSeek is maybe being added to that category."

Related article

Comments (50)