Anthropic Test Reveals DeepSeek AI's Potential Threat in AI Safety

In recent evaluations, the Chinese AI company DeepSeek has shown remarkable performance and cost-efficiency, leading to significant market disruptions such as the declining shares of prominent tech firms like NVIDIA. Despite these achievements, concerns have been raised regarding the potential risks DeepSeek AI poses to national safety. Experts from Anthropic, known for their work with Claude AI, have identified severe vulnerabilities in DeepSeek's ability to handle harmful prompts.

Examination Highlights DeepSeek AI's Safety Flaws

Anthropic, an industry leader poised to enhance Amazon's future AI-driven Alexa, regularly scrutinizes various AI models for their susceptibility to 'jailbreaking'—a method of bypassing built-in safeguards to produce dangerous content. According to Dario Amodei, Anthropic's CEO, DeepSeek AI fails to effectively restrict access to sensitive information, including data concerning biological weapons, labeling its performance in security as significantly lacking compared to other models tested.

Amodei pointed out that DeepSeek AI models, specifically their reasoning-focused R1 model, exhibit no preventive measures against disseminating rare and potentially harmful data. This information is elusive, as it cannot be found through conventional sources like Google or standard textbooks.

Cisco's Analysis Confirms Worrying Trends

Corroborating Anthropic's findings, Cisco's own tests revealed alarming results. DeepSeek's R1 model displayed a 100% Attack Success Rate (ASR), indicating total failure in thwarting harmful prompt generation. These tests also targeted other AI frameworks, resulting in a concerning 86% ASR for the GPT 1.5 Pro model and a 96% ASR for Meta’s Llama 3.1 405B, highlighting pervasive security gaps across the industry.

Though Amodei refrains from labeling DeepSeek models as inherently hazardous, he emphasizes the urgency for DeepSeek's developers to address these AI safety issues earnestly. He also recognizes DeepSeek's potential as a formidable player in the AI sector, underscoring the importance of prioritizing security enhancements.