OpenAI and Google Sound Alarm as Chinese Firms Clone AI Models Through Mass Prompting Attacks
Summary
OpenAI and Google report that Chinese AI companies, including DeepSeek, are cloning their advanced AI models through mass prompting attacks that bypass restrictions, creating potentially dangerous copies without safety guardrails.
Key Points
- OpenAI warns US lawmakers that Chinese AI firm DeepSeek is using distillation techniques to 'free-ride' on OpenAI's model capabilities through obfuscated methods that circumvent access restrictions
- Google's Threat Intelligence Group reports a surge of 'distillation attacks' where commercially motivated actors prompt Gemini thousands of times to learn how it works and clone the model
- Both companies express concern that cloned models lack essential safety guardrails, potentially resulting in dangerous outputs in high-risk domains despite distillation being a commonly used AI training technique