Anthropic CEO says DeepSeek was ‘the worst’ on a critical bioweapons data safety test

Anthropic’s CEO Dario Amodei is fearful about competitor DeepSeek, the Chinese language AI firm that took Silicon Valley by storm with its R1 model. And his issues may very well be extra critical than the everyday ones raised about DeepSeek sending consumer information again to China. 

In an interview on Jordan Schneider’s ChinaTalk podcast, Amodei mentioned DeepSeek generated uncommon details about bioweapons in a security take a look at run by Anthropic.

DeepSeek’s efficiency was “the worst of mainly any mannequin we’d ever examined,” Amodei claimed. “It had completely no blocks in any way towards producing this info.”

Amodei acknowledged that this was a part of evaluations Anthropic routinely runs on numerous AI fashions to evaluate their potential nationwide safety dangers. His group appears to be like at whether or not fashions can generate bioweapons-related info that isn’t simply discovered on Google or in textbooks. Anthropic positions itself because the AI foundational mannequin supplier that takes safety seriously.

Amodei mentioned he didn’t assume DeepSeek’s fashions in the present day are “actually harmful” in offering uncommon and harmful info however that they could be within the close to future. Though he praised DeepSeek’s group as “gifted engineers,” he suggested the corporate to “take severely these AI security concerns.”

Amodei has additionally supported strong export controls on chips to China, citing issues that they might give China’s army an edge.

Amodei didn’t make clear within the ChinaTalk interview which DeepSeek mannequin Anthropic examined, nor did he give extra technical particulars about these exams. Anthropic didn’t instantly reply to a request for remark from TechCrunch. Neither did DeepSeek.

DeepSeek’s rise has sparked issues about its security elsewhere, too. For instance, Cisco safety researchers said last week that DeepSeek R1 failed to dam any dangerous prompts in its security exams, reaching a 100% jailbreak success fee.

Cisco didn’t point out bioweapons however mentioned it was capable of get DeepSeek to generate dangerous details about cybercrime and different unlawful actions. It’s price mentioning, although, that Meta’s Llama-3.1-405B and OpenAI’s GPT-4o additionally had excessive failure charges of 96% and 86%, respectively. 

It stays to be seen whether or not security issues like these will make a critical dent in DeepSeek’s rapid adoption. Firms like AWS and Microsoft have publicly touted integrating R1 into their cloud platforms — sarcastically sufficient, provided that Amazon is Anthropic’s biggest investor.

Alternatively, there’s a growing list of nations, companies, and particularly authorities organizations just like the U.S. Navy and the Pentagon which have began banning DeepSeek. 

Time will inform if these efforts catch on or if DeepSeek’s international rise will proceed. Both manner, Amodei says he does take into account DeepSeek a brand new competitor that’s on the extent of the U.S.’s prime AI firms.

“The brand new truth right here is that there’s a brand new competitor,” he mentioned on ChinaTalk. “Within the large firms that may practice AI — Anthropic, OpenAI, Google, maybe Meta and xAI — now DeepSeek is possibly being added to that class.”