AI models expose personal information vulnerabilities, researchers warn

In recent months, researchers from the University of Tübingen, the Max Planck Institute, the AI safety institute MATS Research, and the security company Snyk have made significant findings regarding the vulnerabilities of AI models. They discovered that certain Chinese AI models may have been trained by distilling reasoning information from leading US models, indicating a potential breach of intellectual property. This research revealed that major AI model providers, including OpenAI, Anthropic, and Google, share a common vulnerability that can lead to personal information leakage, such as passwords and API keys. Although this specific vulnerability has been addressed, the researchers demonstrated that their method could still uncover reasoning traces that may contain sensitive information. The implications of this research are profound, as it raises concerns about the security of AI systems and the potential for large-scale reasoning distillation attacks. The researchers also noted that the open-weight Chinese model Kimi K3 produced outputs strikingly similar to the reasoning traces of Claude Opus 4.8 and GPT 5.6 Sol, suggesting that distillation techniques are being employed to replicate the capabilities of advanced models. However, they emphasized that their findings do not conclusively establish a causal link between the models, leaving room for further investigation into the practices of AI development in different countries. The ongoing debate surrounding distillation techniques highlights the ethical considerations and potential risks associated with AI advancements, particularly in the context of international competition and security.