The Informed Layer · Newsscan
Daily briefing

Anthropic builds wet lab; MCP adoption expands

← All issues

Anthropic / Claude ecosystem

Anthropic Quietly Establishes Physical Biology Lab for Drug Discovery

Anthropic has set up a physical wet-lab facility in the Bay Area to accelerate AI-directed drug discovery, marking an expansion of Claude into experimental biological validation beyond pure in silico modeling.

Anthropic Uses Claude to Optimize 30+ Open-Source Biomolecular Models

Anthropic published research showing Claude optimized over 30 biomolecular foundation models, achieving up to 4x GPU inference speedups and enabling single-node processing of sequences exceeding 10,000 tokens through custom FlashPairformer kernels.

Anthropic Discloses Claude Now Leads 26% of Its Own Model R&D

Anthropic disclosed internal governance metrics revealing that Claude models autonomously lead 26% of their own architectural R&D workflows and assist with 90%, calling on frontier labs to publish standardized metrics on recursive self-improvement.

Frontier model providers

OpenAI Launches Astra for Law Paired with 230M-URL Case Law Index

OpenAI introduced Astra for Law, a specialized configuration of GPT-6 Astra tailored for legal analysis and integrated with a proprietary search index covering 230 million US legal URLs and 26 industry partner connectors.

Cohere Launches Model Vault Encrypted Tier for Hardware-Isolated Inference

Cohere introduced Model Vault Encrypted, providing enterprise inference hosted inside attested confidential computing enclaves that cryptographically prevent the model host from accessing client prompt contexts or generated outputs.

AI developer tooling & infrastructure

Plugin4Shell Zero-Click RCE Flaw Disclosed Across AI Coding Agents

Security researchers revealed 'Plugin4Shell', a critical supply-chain flaw affecting major AI coding assistants including Claude Code, OpenAI Codex, GitHub Copilot, and Gemini CLI, allowing malicious repository owners to bypass SHA pinning and achieve remote code execution.

Oracle Releases SQLcl MCP Server Architecture for Enterprise Database Agents

Oracle published technical integration patterns connecting AI agents to Oracle AI Databases using the Model Context Protocol via SQLcl, providing bounded execution environments and structured enterprise data access.

Relativity aiR Connects to ChatGPT Enterprise via Model Context Protocol

Legal technology provider Relativity integrated its aiR platform with ChatGPT Enterprise via MCP, enabling legal teams to run e-discovery and compliance workflows directly from OpenAI's enterprise chat interface.

GitLab 19.4 Embeds Agentic Automation and MCP Server Tools

GitLab launched version 19.4 featuring native Model Context Protocol server tools in public beta alongside agentic DevSecOps automation for merge requests, work tracking, and vulnerability remediation across open and closed models.

Cloudflare Open-Sources Sandboxed AI Security Audit Agent Skill

Cloudflare released an open-source AI security audit skill on GitHub designed to allow autonomous AI agents to conduct systematic vulnerability scanning and remediation within bounded execution sandboxes.

Cloud & platform providers

Moonshot AI Deploys 2.8T Parameter Kimi K3 on Amazon Bedrock

Moonshot AI made its 2.8-trillion parameter open-weight model Kimi K3 available on Amazon Bedrock, featuring a 1-million-token context window and explicit prompt caching capabilities.

AWS Introduces SageMaker HyperPod Inference Gateway for GPU Routing

AWS unveiled the SageMaker HyperPod Inference Gateway, a Kubernetes-native request router that monitors real-time GPU load to reduce time-to-first-token latency by up to 82% across large inference clusters.

Huawei Cloud Unveils Full-Stack Enterprise Agentic AI Infrastructure

Huawei Cloud rolled out a comprehensive agentic cloud suite including AI Cluster Service (AICS), Context Memory Storage, Agentic MaaS, and specialized industry foundation hubs to support enterprise agent deployment.

Google Home Opens Smart Home Platform to Third-Party AI Agents via MCP

Google enabled Model Context Protocol support across Google Home, making it the first major smart home ecosystem to allow external AI agents such as Claude to query and control connected home IoT hardware.

AI policy, regulation & governance

Google Expands AI & Economy Research Bench with Leading Economists

Google expanded its dedicated AI & Economy research group to collaborate with Nobel laureates and academic economists, aiming to systematically measure macroeconomic productivity, displacement, and labor impacts driven by frontier AI adoption.

Meta Oversight Board Issues Warning Over Deepfake Hate Speech Inaction

Meta's Oversight Board ruled that Meta erred in failing to remove an AI-generated synthetic video featuring hate speech against refugees, finding the company's existing deepfake moderation, labeling, and synthetic media policies structurally insufficient.

European AI Startups Reject US Calls for Frontier AI Development Slowdown

Prominent European AI founders and organizations publicly pushed back against US frontier lab proposals for voluntary AI development pacing, arguing that slowdown rhetoric risks cementing US market dominance and stalling European competitiveness.

xAI Explores Acquiring Training Data from Defunct Startups

xAI and SpaceX are exploring buying proprietary datasets and user histories from failed startups to train upcoming Grok architectures, drawing heightened scrutiny from European privacy regulators over data provenance.

Ninth Circuit Rejects Coders' DMCA Claims Against GitHub Copilot and OpenAI

The US Ninth Circuit Court of Appeals ruled that AI coding tools like GitHub Copilot generate net-new synthetic output rather than stripping copyright management information under DMCA Section 1202(b), while allowing breach-of-contract claims to proceed.

California Governor Issues Executive Order Advancing Frontier AI Kill Switch Mechanisms

California Governor Gavin Newsom signed an executive order directing state agencies and expert panels to develop mandatory safety guardrails for frontier AI models, including external evaluations and verified emergency shutdown 'kill switch' protocols.

Australian Home Affairs Explores Mandatory Breach Reporting for AI Agent Cyber Attacks

Australia's Department of Home Affairs is weighing amendments to the SOCI Act to enforce mandatory breach notification across all corporate sectors whenever systems are compromised by autonomous AI agents or AI-driven tooling.

Anthropic joined OpenAI in pausing domestic AI model training in Australia as tech leaders and the federal government remain deadlocked following the government's abandonment of proposed AI copyright exemptions.

ASD Cyber Chief Backs Domestic Model Training with Anthropic and OpenAI for Defence

The head of the Australian Signals Directorate stated that Australia must train AI models domestically for national security resilience, endorsing Anthropic and OpenAI as preferred enterprise partners despite ongoing domestic copyright tensions.

Industry & market moves

AWS and NVIDIA Partner on 2 Million Next-Gen GPU Fleet and Federal AI Factories

AWS and NVIDIA announced a massive expansion of their cloud partnership to deploy 2 million Blackwell Ultra and Rubin GPUs alongside Vera CPUs by 2028, including dedicated 100,000-GPU air-gapped federal AI factories for classified defense workloads.

Chinese AI Agent Startup Manus Pursues $500M at $4B Valuation After Blocked Meta Deal

Autonomous AI agent startup Manus is raising $500 million at a $4 billion valuation to resume independent operations after Chinese regulatory intervention halted a proposed $2 billion buyout by Meta.

SoftBank Agrees to Acquire Hyundai's Robotics and AI Institute

SoftBank reached an agreement to acquire the Robotics and AI Institute from Hyundai, reuniting Boston Dynamics founder Marc Raibert with SoftBank's growing industrial and physical AI ecosystem.

Raindrop Raises $50M Series A for Production AI Agent Failure Detection

AI observability startup Raindrop secured $50 million in Series A funding led by CRV to expand its anomaly detection and pre-deployment simulation platform that identifies silent logic failures in production AI agents.

Unstop Acquires PerspectAI to Integrate AI Psychometrics and Game Science

Talent platform Unstop acquired PerspectAI to incorporate AI-driven game science and behavioral psychometric assessment models across its 32-million-user recruitment platform.

China's D-Robotics Secures $400M Series C for Embodied AI Silicon Platform

Embodied AI chipmaker D-Robotics raised $400 million in Series C financing to scale production of dedicated robotics processors designed for humanoid manipulation and autonomous edge perception.

AI product & feature launches

Cognition Launches Devin Code Scans for Automated Repository Maintenance

Cognition released Code Scans for its Devin software engineering agent, using a distributed MapReduce architecture to autonomously analyze entire codebases against performance, security, and refactoring goals before submitting completed PRs.

Samsung Medical Center Demos Multi-Humanoid Surgical Assistant Team

Samsung Medical Center completed a surgical demonstration where multiple humanoid robots collaborated directly with a human surgeon, performing dynamic tasks including instrument passing, endoscopic tracking, and tissue traction.

Xenon Open-Sources 397B Parameter Hunmin VLM Computer-Operating Model

Korean AI lab Xenon released weights for Hunmin VLM 397B, an open-weight vision-language model optimized specifically for GUI perception, screen parsing, and autonomous OS interaction.

Unbiased.ai Unveils Stealth Reasoning Model Union Alpha

AI startup unbiased.ai introduced Union Alpha, a high-performance reasoning model that matched GPT-6 Astra and Claude Fable 5.1 on standardized benchmark suites during stealth evaluation before transitioning to paid commercial API access.

Meta Launches Native macOS Desktop Application for Muse AI Agent

Meta released a native macOS application for its Muse personal agent, enabling local OS integration to autonomously read and interact with desktop files, messages, calendars, and mail client workflows.

Research with immediate practical relevance

Google DeepMind Demonstrates Dream-RSI Cutting Search Calls by 162x

DeepMind published Dream-RSI, a framework showing recursive self-improvement through learned meta-exploration strategies that reduces requisite search calls by up to 162 times during autonomous problem solving.

Google Research Introduces Generative UI Framework for Educational Simulations

Google Research released a generative UI system enabling educators to generate customized, interactive STEM simulations and learning modules, publishing an initial library of over 30 peer-reviewed interactive tools.

Figure AI Demonstrates Zero-Shot Humanoid Generalization Across 30 Unseen Homes

Figure AI published benchmark data for its Helix 2.5 model, achieving a 56% task completion rate across 30 novel domestic environments without prior room-specific training data or fine-tuning.

Stanford Virtual Lab of 37,000 AI Scientist Agents Designs Validated Cancer Therapy

Researchers at Stanford Medicine published findings in Science detailing a virtual biotechnology organization staffed by 37,000 autonomous AI agents that designed a lung cancer antibody-drug conjugate which achieved independent laboratory validation and FDA designation.