Fact Checked by 
DeepSeek, the China-based AI model that has made waves in the past few weeks, has been accused of generating rare information about bioweapons. The Wall Street Journal reported that the app is able to produce potentially hazardous materials, unlike some of its counterparts.
Testing done by Journal experts found that DeepSeek provided instructions for a bioweapon attack, and could also craft a phishing email with a malware code. Using a jailbreak, AI safety experts discovered that it could also generate instructions to modify bird flu, as well as create a social media campaign about self-harm among teens. Some online analysts claimed that they were able to jailbreak DeepSeek into generating recipes for Quaaludes and methamphetamines.
Developers from rival firms, such as OpenAI and Meta, have already introduced safeguards to prevent their technology from being tricked into making illicit responses. These tactics are called jailbreaking.
‘No guardrails’ on DeepSeek
However, in the past week, Anthropic’s CEO, Dario Amodei, echoed the findings. Speaking on Jordan Schneider’s ChinaTalk podcast, he said that Anthropic analysts could make DeepSeek produce information about bioweapons. He said that DeepSeek’s technology was “the worst of basically any model we’d ever tested,” adding that it “had absolutely no blocks whatsoever against generating this information.”
My thoughts on China, export controls and two possible futures https://t.co/LZAANEcZJW
— Dario Amodei (@DarioAmodei) January 29, 2025
Anthropic, a company focused on AI safety research, regularly tests AI models to identify possible risks. Amodei shared his thoughts on DeepSeek, saying he doesn’t believe its models are currently dangerous when it comes to generating rare and harmful information—but that could change soon. While he acknowledged DeepSeek’s team as “talented engineers,” he also urged them to “take seriously these AI safety considerations.”
In a blog post published in January, Amodei added: “Combined with its large industrial base and military-strategic advantages, this could help China take a commanding lead on the global stage, not just for AI but for everything.”
DeepSeek has risen in popularity when it surpassed ChatGPT in downloads on Apple’s App Store at the end of last month. The R1 model has been highly praised for its efficiency and ability to compete with the $200 a month o1 model from OpenAI, maker of ChatGPT.
Featured image: DeepSeek / Canva
