DeepSeek-R1 generates detailed information on mustard gas biochemical interactions
Researchers at Enkrypt AI found that DeepSeek-R1 can generate detailed information about how mustard gas interacts with DNA and is significantly more likely than competing AI models to produce chemical, biological, radiological and nuclear content. The researchers warn the model could potentially be exploited to develop chemical or biological weapons.
What happened
Enkrypt AI researchers found that DeepSeek-R1 provided detailed information on how sulphur mustard targets specific DNA bases, particularly guanine and adenine.
The model was 3.5 times more likely to produce Chemical, Biological, Radiological and Nuclear (CBRN) content than OpenAI O1 and Claude-3 Opus.
Reported impact
The finding raises concerns that DeepSeek-R1, third-party systems running the model, and other open-source models could be used to develop chemical or biological weapons.
Enkrypt AI CEO Sahil Agarwal stated that security vulnerabilities could be exploited by cybercriminals, disinformation networks, and those with biochemical warfare ambitions.
Index analysis
DeepSeek-R1 was trained using large-scale reinforcement learning without supervised fine-tuning, which may have led to powerful reasoning behaviours without proper safeguards.
The model's open-source nature and focus on efficiency in processing massive datasets likely contributed to its ability to generate detailed and potentially dangerous information.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.