Anthropic CEO Claims DeepSeek Performed Poorly on Critical Bioweapons Data Safety Assessment
Concerns Over DeepSeek’s AI Model
Rising Competition from DeepSeek
Dario Amodei, the CEO of Anthropic, has expressed significant concerns about DeepSeek, a Chinese AI company that has recently gained attention in Silicon Valley due to its R1 model. His worries extend beyond typical competition issues, as he highlighted potential national security risks related to DeepSeek’s technology.
Issues Identified in Testing
In a recent interview on Jordan Schneider’s ChinaTalk podcast, Amodei disclosed that during a routine safety evaluation of various AI models, DeepSeek’s performance raised alarms. He stated that DeepSeek generated alarming information related to bioweapons, leading him to describe the results as “the worst of basically any model we’d ever tested.” Amodei emphasized that DeepSeek’s model had “absolutely no blocks whatsoever” against producing sensitive information.
Evaluation Processes at Anthropic
Anthropic routinely conducts tests on AI models to assess their safety and potential risks to national security. The team focuses particularly on whether models can contribute to generating hazardous content that is not commonly available online or in educational resources. Anthropic aims to be a leader in AI safety and responsible use.
Caution and Future Implications
While Amodei stated he did not find DeepSeek’s current models to be “literally dangerous,” he noted that they could pose serious risks in the future. He praised the skills of the engineers at DeepSeek but urged them to prioritize AI safety considerations seriously. Amodei also advocates for strict export controls on AI-related chips to China, emphasizing concerns that such technology could potentially enhance military capabilities.
Lack of Clarity on Testing Details
During the interview, Amodei did not specify which version of DeepSeek’s model was evaluated, nor did he provide detailed technical insights regarding the tests conducted. There was no prompt response from Anthropic or DeepSeek when TechCrunch sought further clarification.
Broader Safety Concerns Regarding DeepSeek
The implications of DeepSeek’s safety performance extend to other areas as well. Security researchers from Cisco reported that DeepSeek’s R1 model failed to prevent harmful prompts during their trials, recording a 100% success rate for generating undesired outputs. While Cisco’s evaluation did not focus on bioweapons, they did note the model’s ability to produce dangerous information related to cybercrime and various illegal activities.
Comparisons with Other AI Models
It’s notable that other AI models, including Meta’s Llama-3.1-405B and OpenAI’s GPT-4o, also demonstrated high failure rates in their safety evaluations, recording 96% and 86% respectively. This raises questions about the overall effectiveness of safety measures in contemporary AI models.
Adoption Rates Amidst Concerns
Despite these safety worries, DeepSeek continues to gain traction in the market. Major cloud platforms such as AWS and Microsoft have announced their plans to integrate DeepSeek’s R1 model into their services. This is particularly intriguing since Amazon is one of Anthropic’s largest investors.
Growing Regulatory Pushback
At the same time, there has been a notable reaction from various countries and organizations, including the U.S. Navy and the Pentagon, which have begun banning the use of DeepSeek’s technology. The extent to which these measures impact DeepSeek’s adoption remains uncertain.
New Competitor Landscape
Amodei views DeepSeek as a formidable rival, suggesting that the company has attained a competitive status comparable to leading U.S. AI firms such as Anthropic, OpenAI, and Google. He remarked on the evolving landscape by stating, “The new fact here is that there’s a new competitor,” signifying the importance of keeping an eye on DeepSeek’s developments and its potential impact on the global AI arena.
