Anthropic governance & safety; Multi-agent AI rises

Anthropic / Claude ecosystem

Inviting hard questions \ Anthropic

Anthropic has launched a public 'hard questions' initiative to transparently address societal concerns about AI governance and benefits. This initiative aims to foster open dialogue and proactive problem-solving regarding the impact of AI.

Anthropic Launches 'Reflection' Tool for Claude AI Users

Anthropic introduces Reflection, a dashboard tool that helps Claude users track and optimize their AI interactions while maintaining privacy and promoting intentional AI usage aligned with the company's constitutional AI principles. It subtly shapes user perception by visualizing usage patterns and suggesting deeper integration.

Anthropic tests a way to delete dangerous knowledge from AI models

Anthropic demonstrates a structural method, called Gradient-Routed Auxiliary Modules (GRAM), to surgically remove dangerous knowledge (e.g., virology, cybersecurity, nuclear physics) from trained AI models without retraining, while holding performance gains at frontier scale.

Introducing Claude apps gateway for AWS

Anthropic has launched the Claude apps gateway for AWS, a self-hosted control plane that allows enterprises to centralize governance over Claude Code and Claude Desktop access, costs, and policies for their teams.

Frontier model providers

Introducing Muse Spark 1.1

Meta has launched Muse Spark 1.1, a multimodal reasoning model that shows major gains in tool and computer use, coding, agentic tasks, and multimodal understanding. It is now available via the Meta Model API public preview.

AI developer tooling & infrastructure

Kastor brings infrastructure-as-code habits to AI agents

Kastor applies infrastructure-as-code patterns, including HCL, plan/apply, diffs, and drift checks, to AI agent development. This enables declarative agent specifications and multi-framework code generation, streamlining the management of AI agent deployments.

Cloud & platform providers

We're rolling out AlphaEvolve widely to solve Google Cloud customers' hardest problems.

Google Cloud is making AlphaEvolve widely available to its enterprise customers, equipping them with advanced AI capabilities to tackle complex computational and data-driven challenges across various industries.

Enhancing enterprise inference on Amazon SageMaker HyperPod with data capture, Hugging Face, NVMe, and Route 53 integration

Amazon SageMaker HyperPod now offers enhanced enterprise inference capabilities, including multi-tier inference data capture, direct Hugging Face model deployment, node-local NVMe cold-start optimization, and automatic DNS management via Route 53.

C4N network and storage-optimized VMs

Google Cloud's C4N instances are now generally available, delivering 33% higher network bandwidth per vCPU and 224% faster packet processing compared to comparable Intel-based offerings, designed for network and storage-intensive workloads.

Microsoft Pilots Agent ROI Dashboard in Foundry, Forms Copilot Hits Global Availability

Microsoft has launched a private preview of its Agent ROI Dashboard in Foundry to quantify AI agent business value and simultaneously rolled out Forms Copilot globally, indicating a shift toward accountable AI spending and accessible AI-driven productivity tools.

Microsoft's Secret Weapon for Cloud Security: AI Agents That Work in Hours, Not Weeks

Microsoft has internally deployed a multi-agent AI system that compresses cloud security reviews from weeks to hours, enabling continuous near-real-time security validation at cloud scale.

Cloudflare Drop: Instant Edge Deploy, No Account | explainx.ai Blog

Cloudflare has launched 'Cloudflare Drop,' a zero-friction service for instant edge deployment: users can drag a folder into their browser to get a live global URL in seconds without needing to sign up, with automatic 60-minute expiry unless claimed to an account.

AI policy, regulation & governance

GhostApproval Flaw Hits Six Major AI Coding Assistants

A new symlink-based vulnerability dubbed 'GhostApproval' (CVE-2026-50549) has been discovered, allowing attackers to bypass approval prompts in AI coding assistants like Amazon Q Developer, Anthropic Claude Code, Augment, Cursor, Google Antigravity, and Windsurf, potentially leading to remote code execution.

MeitY to discuss separate AI regulation law: S Krishnan

India's Ministry of Electronics and Information Technology (MeitY) has initiated a formal stakeholder consultation process for a dedicated AI regulation law, reversing its earlier stance against new AI-specific legislation.

China’s financial regulator draws up an AI safety rulebook for banks

China's National Financial Regulatory Authority (NFRA) has published its first formal AI safety guidance, designating high-risk AI applications in the banking and insurance sectors, including fund transfers, credit approval, and underwriting, that require stringent oversight.

AI's opportunities and risks front and centre on Albanese government's agenda

The Australian government, under Prime Minister Albanese, is placing AI policy and regulation at the forefront of its agenda, driven by weak economic growth and concerns about AI's impact on the labor market.

Industry & market moves

HUMAIN and Cohere Launch Strategic Sovereign AI Compute Collaboration

HUMAIN and Cohere have launched a strategic collaboration for sovereign AI compute, establishing Cohere's first significant international AI compute deployment outside North America with 50MW of dedicated Saudi infrastructure, scaling over five years for frontier model development and sovereign AI initiatives.

AI unicorn Mercor acquires Deeptune after founder Brendan Foody backed the startup

AI unicorn Mercor, founded by Brendan Foody, has acquired Deeptune after having quietly invested in the startup's Series A round, signaling consolidation in the AI training infrastructure market.

Mukesh Bansal-led Nurix AI acquires Verloop to expand enterprise offerings

Nurix AI has acquired Verloop to expand its conversational AI platform, integrating Verloop's customer base of over 20 million with Nurix's enterprise automation software across voice and chat channels.

Reevo Strengthens Its AI-Powered Prospecting and Outreach With Acquisition of Ciro

Reevo has acquired Ciro, a B2B prospecting agent, to integrate AI-powered prospecting directly into its Revenue Operating System platform, aiming to streamline sales workflows and eliminate tool sprawl.

AI product & feature launches

Robbyant Unveils LingBot-World 2.0: Pioneering Hour-Long Real-Time Generation in World Models

Robbyant has unveiled LingBot-World 2.0 (Infinity), which achieves hour-long continuous real-time generation with zero quality drift in world models. This represents a significant advancement in interactive embodied AI.

NVIDIA Releases Nemotron-Labs-3-Puzzle-75B-A9B: A Compressed Hybrid MoE LLM Delivering 2.03x Server Throughput at Matched User Throughput

NVIDIA has released Nemotron-Labs-3-Puzzle-75B-A9B, a new compressed hybrid Mixture-of-Experts (MoE) LLM. It delivers 2.03x server throughput on decode-heavy workloads and enables 8x concurrent 1M-token requests on a single H100 GPU.

Burn2kicad: AI Agents Cross the Chasm from Code to Circuit Board Design

The burn2kicad open-source project has released its v0.2.0, which includes the first widely accessible AI agent that translates natural language directly into manufacturable PCB designs. This bridges the digital-physical divide for hobbyists and software engineers.

Research with immediate practical relevance

Open-Source World Model AlayaWorld Sustains Interactive Play Past the One-Minute Mark

Shanda AI Research's ALAYA Lab has released AlayaWorld, an open-source world model framework, that can sustain interactive, spatially coherent video generation for over 60 seconds of continuous play, representing a significant advance in training embodied AI systems.

Robbyant Releases LingBot-VLA 2.0: An Open-Source 6B Vision-Language-Action (VLA) Model for Cross-Embodiment Robot Manipulation

Robbyant (Ant Group) has released LingBot-VLA 2.0, an open-source 6B vision-language-action model for cross-embodiment robot manipulation. It was trained on 60,000 hours of data across 20 robot configurations, improving generalization and expanding its action space.

Multi-Agent AI Safety Cannot Be Fixed by Better Models Alone, Study Shows

New research by Yujiao Chen, presented in the Institutional Red-Teaming framework paper, suggests that multi-agent AI safety depends more on deployment governance rules and institutional constraints than on model alignment alone. It indicates that prompt-level safety rules fail where structural enforcement succeeds.