\n<\/p>\n
Anthropic\u2019s CEO Dario Amodei is worried about competitor DeepSeek, the Chinese AI company that took Silicon Valley by storm<\/a> with its R1 model<\/a>. And his concerns could be more serious than the typical ones raised about DeepSeek sending user data back to China.\u00a0<\/p>\n In an interview<\/a> on Jordan Schneider\u2019s ChinaTalk podcast, Amodei said DeepSeek generated rare information about bioweapons in a safety test run by Anthropic.<\/p>\n DeepSeek\u2019s performance was \u201cthe worst of basically any model we\u2019d ever tested,\u201d Amodei claimed. \u201cIt had absolutely no blocks whatsoever against generating this information.\u201d <\/p>\n Amodei stated that this was part of evaluations Anthropic routinely runs on various AI models to assess their potential national security risks. His team looks at whether models can generate bioweapons-related information that isn\u2019t easily found on Google or in textbooks. Anthropic positions itself as the AI foundational model provider that takes safety seriously.<\/a><\/p>\n Amodei said he didn\u2019t think DeepSeek\u2019s models today are \u201cliterally dangerous\u201d in providing rare and dangerous information but that they might be in the near future. Although he praised DeepSeek\u2019s team as \u201ctalented engineers,\u201d he advised the company to \u201ctake seriously these AI safety considerations.\u201d<\/p>\n Amodei has also supported strong export controls<\/a> on chips to China, citing concerns that they could give China\u2019s military an edge.<\/p>\n Amodei didn\u2019t clarify in the ChinaTalk interview which DeepSeek model Anthropic tested, nor did he give more technical details about these tests. Anthropic didn\u2019t immediately reply to a request for comment from TechCrunch. Neither did DeepSeek.<\/p>\n DeepSeek\u2019s rise has sparked concerns about its safety elsewhere, too. For example, Cisco security researchers said last week<\/a> that DeepSeek R1 failed to block any harmful prompts in its safety tests, achieving a 100% jailbreak success rate.<\/p>\n Cisco didn\u2019t mention bioweapons but said it was able to get DeepSeek to generate harmful information about cybercrime and other illegal activities. It\u2019s worth mentioning, though, that Meta\u2019s Llama-3.1-405B and OpenAI\u2019s GPT-4o also had high failure rates of 96% and 86%, respectively.\u00a0<\/p>\n It remains to be seen whether safety concerns like these will make a serious dent in DeepSeek\u2019s rapid adoption<\/a>. Companies like AWS<\/a> and Microsoft<\/a> have publicly touted integrating R1 into their cloud platforms \u2014 ironically enough, given that Amazon is Anthropic\u2019s biggest investor<\/a>.<\/p>\n On the other hand, there\u2019s a growing list<\/a> of countries, companies<\/a>, and especially government organizations like the U.S. Navy and the Pentagon<\/a> that have started banning DeepSeek.\u00a0<\/p>\n Time will tell if these efforts catch on or if DeepSeek\u2019s global rise will continue. Either way, Amodei says he does consider DeepSeek a new competitor that\u2019s on the level of the U.S.\u2019s top AI companies.<\/p>\n \u201cThe new fact here is that there\u2019s a new competitor,\u201d he said on ChinaTalk. \u201cIn the big companies that can train AI \u2014 Anthropic, OpenAI, Google, perhaps Meta and xAI \u2014 now DeepSeek is maybe being added to that category.\u201d<\/p>\n<\/div>\n