Anthropic CEO Dario Amodei has raised serious concerns regarding DeepSeek, a Chinese AI company that has gained significant attention in Silicon Valley, particularly with its R1 model. While much of the discourse surrounding DeepSeek has centered on data privacy and its ties to China, Amodei’s warnings focus on an even more alarming issue: the model’s performance in critical safety tests related to bioweapons information.
In a recent interview on Jordan Schneider’s ChinaTalk podcast, Amodei disclosed that DeepSeek’s model performed “the worst of basically any model we’d ever tested” during a bioweapons data safety evaluation conducted by Anthropic. According to Amodei, DeepSeek showed “absolutely no blocks whatsoever against generating this information,” which could potentially pose significant national security risks.
Anthropic regularly evaluates AI models to assess their potential for misuse, particularly in generating sensitive or dangerous content that may not be readily accessible through common sources such as Google or textbooks. The company prides itself on being a leader in AI safety, emphasizing the need to protect foundational models from misuse.
Amodei clarified that while DeepSeek’s current models may not be “literally dangerous,” they could become so in the future if these vulnerabilities are not addressed. Although he praised DeepSeek’s team as “talented engineers,” he stressed the importance of prioritizing AI safety considerations.
This is not the first time DeepSeek has faced criticism for its safety protocols. In a separate report, Cisco security researchers noted that DeepSeek R1 failed to block harmful prompts during their tests, achieving a concerning 100% jailbreak success rate. While Cisco’s tests focused on cybercrime and illegal activities rather than bioweapons, their findings align with Amodei’s concerns. Notably, even leading models like Meta’s Llama-3.1-405B and OpenAI’s GPT-4 have shown vulnerabilities, with failure rates of 96% and 86%, respectively.
Despite these alarming findings, DeepSeek continues to gain traction. Major tech companies, such as AWS and Microsoft, have integrated DeepSeek R1 into their cloud platforms—despite the fact that Amazon is Anthropic’s largest investor. Some organizations, however, are adopting a more cautious stance. For instance, government entities such as the U.S. Navy and the Pentagon have begun banning DeepSeek due to these safety concerns.
Amodei’s comments highlight the growing recognition of DeepSeek as a formidable competitor in the global AI landscape. “The new fact here is that there’s a new competitor,” he stated 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.”
As the debate over AI safety intensifies, it remains uncertain whether these concerns will slow DeepSeek’s rapid adoption or if the company can resolve these vulnerabilities before they escalate further. One thing is clear: the rise of DeepSeek signifies a pivotal moment in the evolving dynamics of AI competition and regulation.
Also Read : Google Aims to Transform Search into an AI-Powered Assistant by 2025
