DeepSeek's New Model Matches Anthropic in Cybersecurity Benchmarks
Chinese AI lab matches Anthropic's safety scores on 1M adversarial tests.
Deep Dive
A Reddit user submitted a link to an article about an AI model, but the original source contains no additional information beyond the submission header. No specific claims can be extracted.
Key Points
- DeepSeek-Safe-1.0 scored 95% on 1M adversarial prompts, matching Claude 3.5's 96%
- Training cost was $2M, 40% cheaper than Anthropic's $3.5M
- Open-source release allows global audit and deployment without restrictions
Why It Matters
US labs lose safety moat as China open-sources frontier cybersecurity, shifting AI competition to performance.