Microsoft launches ExCyTIn-Bench, an open-source benchmark that tests how AI agents handle real-world cybersecurity investigations inside a simulated Azure SOC using 57 Sentinel log tables.

Shadow AI on the rise: Microsoft research finds 71% of UK employees use unapproved consumer AI tools at work and 51% do so weekly, creating serious risks around data privacy, compliance, and cybersecurity.

Autonomous vulnerability repair: Google’s CodeMender is an AI-powered agent that automatically finds, patches, and rewrites insecure code — fixing 72 vulnerabilities across major open-source projects while eliminating entire classes of security flaws. Source @

Guardrails bypassed: Researchers at HiddenLayer discovered that OpenAI’s new Guardrails safety framework, launched with AgentKit, can be easily bypassed — showing that when the same model acts as both responder and safety checker, it can be fooled by jailbreaks and prompt injections. Source @ HackRead

A consortium (BlackRock, Microsoft, Nvidia, xAI, MGX, GIP, and Temasek) announced a $40B acquisition of Aligned Data Centers, forming the first major deal under the AI Infrastructure Partnership to scale AI-optimized data centers. Source @ CNBC

OpenAI and Broadcom revealed a multi-year collaboration to co-develop custom AI accelerators and networking systems, aiming for deployment of 10 GW of hardware by 2029. OpenAI

Oracle launched OCI Zettascale10, a new cloud supercluster architecture that connects massive numbers of NVIDIA GPUs across data centers to deliver multi-gigawatt AI performance and improved utilization. Oracle

Nscale struck a major infrastructure deal with Microsoft, contracting ~200,000 NVIDIA GB300 GPUs across Europe and the U.S. to expand hyperscale AI capacity in partnership with Dell. Nscale+1

At the 2025 Open Compute Project Global Summit, AMD emphasized that the future of AI infrastructure will be built on open standards and collaboration — showcasing the new Helios rack architecture and leadership in UALink™, UltraEthernet, and E-SUN initiative. AMD

Apple’s new M5 chip delivers over 4× the AI GPU performance of M4, powered by a next-generation 10-core GPU with a Neural Accelerator in every core, a faster CPU and Neural Engine, and 30% higher unified memory bandwidth (153 GB/s). Apple

Microsoft introduced Agent Mode in Office apps and the new Office Agent in Copilot chat, marking a major step toward agentic productivity in the workplace.

OpenAI Agent Builder uses a node-based interface for building agent workflows, which is useful for quickly creating and visualizing the steps of an agent.

Salesforce unveiled Agentforce 360, the world’s first platform connecting humans and AI agents in one trusted system, ushering in the era of the Agentic Enterprise, Salesforce.

Oracle introduced AI agents embedded within Fusion Cloud Applications, enabling smarter, faster decisions across finance, HR, supply chain, sales, marketing, and service.

Strengthen AI Trust and Security: Build governance frameworks that balance innovation with safety — auditing AI use, securing data pipelines, and adopting trusted benchmarks to validate model behavior.

Invest in Scalable, Open Infrastructure: Modernize enterprise systems for multi-cloud, GPU-optimized, and open-standard architectures that can support agentic AI workloads at scale.

Empower Human-Agent Collaboration: Shift from tools to teamwork — integrate AI agents across productivity, CRM, and ERP platforms to elevate decision-making, creativity, and business execution.