Australia investigates OpenAI breach of government health website

Google DeepMind has introduced private, server-side memory capabilities for its Private AI Compute offering, designed to enable personal AI applications while maintaining data privacy. The advancement allows AI models to access and utilize memory on secure servers without exposing user data to the broader system. This development addresses a key technical challenge in deploying private AI systems that require persistent context while maintaining cryptographic isolation.

Chinese AI researchers are increasingly choosing to remain and work in China rather than relocate abroad, according to a Carnegie China study. The share of top AI researchers working in China has risen from 27.1%, marking a significant shift in the global distribution of elite AI talent. This trend reflects both improved opportunities within China's AI ecosystem and changing career preferences among Chinese researchers.

China's internet regulator is investigating DeepSeek and Moonshot AI following allegations by Anthropic that both companies routed sensitive user data to Claude models without authorization. Anthropic published a 154-page report on September 10 detailing how seven Chinese companies were using Claude illicitly at scale, including an example where DeepSeek relayed requests from engineers building a police surveillance system to Claude. The investigation marks a significant escalation in scrutiny of data practices among Chinese AI firms and raises questions about the security of proprietary AI systems.

AI security requires engineering discipline across the full agent stack, from models through runtime environments, with enforceable controls at each layer rather than relying on agent reasoning alone. Saša Zdjelar argues that organizations must apply established security principles to new AI operating conditions, implement traceable identities and bounded permissions, and gather evidence that protections work before deployment. NVIDIA's OpenShell and partner tools like Cisco's DefenseClaw demonstrate how to enforce policies outside an agent's reach.