AI could eat itself: Competitors (..) steal their secrets and clone them (www.theregister.com)

Google and OpenAI have raised alarms about a growing threat from competitors, specifically China's DeepSeek, which are attempting to steal the intellectual property of leading AI models through a process called "distillation attacks." This method involves probing models with strategic prompts to uncover their underlying reasoning, enabling rivals to replicate sophisticated capabilities. Google reported that campaigns have employed over 100,000 prompts to imitate their model, Gemini, reflecting a significant risk for AI companies that invest heavily in developing their technology. Both firms have acknowledged the difficulty in completely preventing such incursions, given the public accessibility of many AI models. The implications of this threat are far-reaching, as it not only challenges the security of proprietary models, but also risks reducing the competitive edge of American AI technologies. As competitor methods have evolved to become more sophisticated—incorporating synthetic data generation and other advanced techniques—OpenAI has called for an "ecosystem security" approach to bolster defenses against unauthorized access. The company advocates for U.S. government involvement in establishing best practices and closing loopholes that allow adversaries like DeepSeek to access American AI models. This situation highlights the ongoing arms race in AI development and the need for collaborative solutions to safeguard against intellectual property theft.