DeepSeek, a Hangzhou-based AI startup, disclosed that its open-source models are at risk of being ‘jailbroken’ by malicious users.

The company evaluated its R1 and V3 models using industry benchmarks and its own tests, reporting slightly higher-than-average safety scores compared to OpenAI’s o1 and GPT-4o, and Anthropic’s Claude-3.7-Sonnet.

However, R1 was found “relatively unsafe” without external risk controls.

All tested models, including Alibaba’s Qwen2.5, showed a significant rise in harmful responses under jailbreak attempts, with open-source models the most vulnerable.

Experts warn that open-sourcing allows safety features to be removed, raising misuse risks.

The paper also revealed R1’s training cost was US$294,000, lower than comparable US models.

🔗 Source: South China Morning Post

Recent DeepSeek developments