Anthropic CEO Dario Amodei has expressed serious concerns over DeepSeek, the Chinese AI company making waves in Silicon Valley with its R1 model. While discussions about data privacy and potential ties to the Chinese government have been common, Amodei’s worries extend beyond that—focusing instead on AI safety risks.
In an interview on the ChinaTalk podcast, Amodei revealed that DeepSeek’s AI model performed poorly in Anthropic’s safety tests, generating rare bioweapons-related information without any safeguards.
💬 “It had absolutely no blocks whatsoever against generating this information. It was the worst of basically any model we’d ever tested.” — Amodei
Anthropic regularly evaluates AI models for national security risks, particularly testing whether they can generate bioweapons-related knowledge that isn’t easily accessible online. While Amodei didn’t disclose which DeepSeek model was tested, he warned that future iterations could pose a greater danger.
DeepSeek’s Security Weaknesses Spark Broader Concerns
Amodei isn’t alone in his concerns. Cisco security researchers recently reported that DeepSeek R1 failed all safety tests, allowing 100% of harmful prompts to bypass restrictions. Their study didn’t mention bioweapons, but they confirmed DeepSeek could generate cybercrime-related information and other illicit content.
For context:
📌 DeepSeek R1’s failure rate: 100%
📌 Meta’s Llama-3.1-405B: 96% failure rate
📌 OpenAI’s GPT-4o: 86% failure rate
These findings have not slowed DeepSeek’s growth, with companies like AWS and Microsoft integrating R1 into their cloud platforms—a notable irony given Amazon’s major investment in Anthropic.
Growing Global Scrutiny on DeepSeek
Amid security concerns, several organizations—including the U.S. Navy and the Pentagon—have started banning DeepSeek’s AI models. Meanwhile, governments and regulatory bodies are weighing their options on whether to impose restrictions on the Chinese AI company.
Despite this, Amodei acknowledges DeepSeek as a serious new competitor in the AI landscape.
💬 “The new fact here is that there’s a new competitor… DeepSeek is maybe being added to the category of companies that can train frontier AI models.”
With AI geopolitics heating up and concerns over AI-generated dangerous content growing, the battle between Anthropic, OpenAI, Google, and now DeepSeek is shaping up to be a defining moment for the future of AI safety and regulation.
