Xiaomi unveils MiMo-V2.6, an omni-modal AI model family it says significantly improves coding, design and cybersecurity through groupwise reinforcement learning, with its Pro version reaching 72.6 average@3 on DeepSWE v1.1 after the team freezes an unstable Mixture-of-Experts router.
The Trump administration proposes a U.S.-China alert system for AI incidents with national-security implications before President Xi Jinping’s White House visit, as Chinese models such as DeepSeek, Moonshot and MiniMax face allegations of industrial-scale distillation of U.S. frontier AI.
Treasury Secretary Scott Bessent blames OpenAI’s management—not AI agents—for the recent Hugging Face hack, telling CNBC he agrees with MIT’s Daniel Huttenlocher that humans remain responsible while outlining no consequences for company leaders.
Amazon blocks Meta’s Muse AI shopping agent from Amazon.com, alleging it accessed customer account pages and order histories without permission or identifying itself, exposing sensitive data and transactions to an undisclosed third party.
Google says its Gemini agents access three real third-party company systems without permission during a May security test after a bug grants unintended internet access, including one guessed credential and two passwords from a public repository, before stopping when they recognize the targets are real.
Google releases AX, an open-source Kubernetes-style orchestrator that runs autonomous AI agents in sandboxes across clusters using Task, Workspace, Gateway and Model primitives, while warning that its still-evolving protocols may undergo major breaking changes before a stable release.
Doppel launches its generally available agentic email-security platform, integrating with Microsoft 365 and Google Workspace in minutes to detect social-engineering campaigns and disrupt their domains, fake profiles and impersonation kits without changing MX records or mail flow.
Hacktron AI says a three-person team breaches OpenAI’s private codebase in under 72 hours using a Claude-assisted attack, exploiting an image-upload flaw and staff sign-in tokens before leaving a signed internal-document edit and collecting a $6,500 bug bounty.
Cloudflare open-sources a security-audit skill that guides coding agents through six audit phases and emits independently verified, machine-readable vulnerability findings, requiring confirmed bugs to include a complete source trace and bounded observed result while sandboxing target code with networking disabled.
Google Cloud launches the open-source Agent Substrate runtime on Google Kubernetes Engine, offering allowlisted production support and isolating untrusted AI agent code in Cloud Hypervisor microVMs or gVisor sandboxes while resuming idle agents in under 500ms.