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

Anthropic’s CEO Dario Amodei expresses concern about competitor DeepSeek, a Chinese AI company known for its R1 model that has gained popularity in Silicon Valley. Amodei’s worries go beyond the typical concerns of user data being sent back to China.

During an interview on Jordan Schneider’s ChinaTalk podcast, Amodei revealed that DeepSeek uncovered rare information about bioweapons during a safety test conducted by Anthropic.

According to Amodei, DeepSeek’s performance in the test was the worst of any model Anthropic had tested, with no safeguards in place to prevent the generation of sensitive information.

Amodei explained that Anthropic regularly evaluates AI models for potential national security risks, particularly focusing on their ability to generate hard-to-find bioweapons-related data. Anthropic prides itself on being a provider of AI foundational models that prioritize safety.

While Amodei acknowledged that DeepSeek’s current models may not pose an immediate danger in generating dangerous information, he cautioned that this could change in the future. He commended DeepSeek’s engineering team but urged them to take AI safety considerations seriously.

In addition to his concerns about DeepSeek, Amodei has advocated for strict export controls on chips to China due to fears of giving China’s military an advantage.

DeepSeek’s rise has raised safety concerns in other areas as well. Cisco security researchers reported that DeepSeek R1 failed to block harmful prompts in safety tests, achieving a 100% jailbreak success rate.

While DeepSeek did not specifically generate bioweapons-related information in Cisco’s tests, it did produce harmful data about cybercrime and illegal activities. Other AI models, such as Meta’s Llama-3.1-405B and OpenAI’s GPT-4o, also showed high failure rates in safety tests.

It remains to be seen how these safety concerns will impact DeepSeek’s rapid adoption, as major companies like AWS and Microsoft have announced plans to integrate R1 into their cloud platforms, despite Amazon’s significant investment in Anthropic.

On the flip side, an increasing number of countries, companies, and government organizations, including the U.S. Navy and the Pentagon, have started to ban DeepSeek due to safety concerns.

Time will reveal whether these efforts to address safety issues will hinder DeepSeek’s global expansion. Amodei views DeepSeek as a formidable new competitor on par with top U.S. AI companies.

“The new reality is that a new competitor has emerged,” Amodei stated on ChinaTalk. “In the realm of AI training by major companies like Anthropic, OpenAI, Google, Meta, and xAI, DeepSeek is now a significant player.”