
DeepSeek R1 Passed the Safety Test Like a Guest Who Skipped the Bouncer Entirely
Ourtake
Reported as 11x more likely than OpenAI's O1 to generate harmful output. The bouncer was apparently never hired.
What happened
Security firm Enkrypt AI tested DeepSeek R1 and reported it was more likely than rival models to produce biased, toxic, harmful and insecure content, including malicious code and extremist recruitment text. The report says the model lacks adequate safety guardrails, and locally installed copies will not receive the app's safety updates.
Sources
- bgr.com · DeepSeek R1 Can Generate More Harmful Content Than Other AI, Study Finds - BGR (opens in a new tab)
- euronews.com · Why DeepSeek's 'major security and safety gaps' are causing concern (opens in a new tab)
- itbrief.asia · DeepSeek-R1 AI model's vulnerabilities pose global threat (opens in a new tab)