Anthropic CEO Dario Amodei has expressed serious concerns about Chinese AI company DeepSeek's performance in recent safety evaluations, revealing that the company's AI model showed concerning behaviors in tests related to bioweapons information.
During an interview on the ChinaTalk podcast, Amodei disclosed that DeepSeek generated sensitive bioweapons-related information during Anthropic's routine safety assessments. He described the model's performance as "the worst of basically any model we'd ever tested," noting that it lacked any protective barriers against producing such dangerous content.
The revelation comes amid growing scrutiny of DeepSeek's safety protocols. Last week, Cisco security researchers reported that DeepSeek R1 failed to block harmful prompts in their tests, achieving a 100% jailbreak success rate. While other AI models like Meta's Llama-3.1-405B and OpenAI's GPT-4 also showed high failure rates, DeepSeek's complete lack of safeguards has raised red flags.
While Amodei acknowledged that DeepSeek's current models aren't "literally dangerous," he warned about potential risks in the near future. He urged the company to prioritize AI safety measures, particularly given their rapid emergence as a major player alongside established companies like OpenAI, Google, and Meta.
The safety concerns have already prompted action from some quarters. The U.S. Navy, Pentagon, and various organizations have begun implementing bans on DeepSeek's technology. However, major cloud providers like AWS and Microsoft continue to integrate DeepSeek's R1 model into their platforms.
Neither Anthropic nor DeepSeek have responded to requests for comment about these safety test results. The specific DeepSeek model tested and detailed technical aspects of the evaluation remain undisclosed.