The Informed Layer · Newsscan
Daily briefing

Agents drive new models; Global AI laws emerge

← All issues

Anthropic / Claude ecosystem

Claude Code 2.1.229 released with SSE Keepalives and hardened Git Push security

Anthropic released Claude Code 2.1.229, which addresses silent stream termination issues in long-reasoning tasks by injecting SSE keepalive pings for Vertex AI and Bedrock. The update also enhances security by locking down git-push auto-approval to prevent the use of dangerous flags like --force and --no-verify.

Anthropic unifies Claude in Chrome browser sessions with Cowork backend for state sync

Anthropic has launched 'Claude in Chrome synced sessions,' enabling the unification of browser sessions with its Cowork backend. This allows Claude's session state, including Skills and Connectors support, to synchronize across multiple client devices.

Frontier model providers

OpenAI launches Computer History, a keylogging system for ChatGPT, replacing screenshot-based Chronicle

OpenAI has introduced Computer History, a new keylogging and event-capture system for ChatGPT that records user interactions locally and sends summaries to OpenAI servers. This replaces the previous screenshot-based Chronicle approach, aiming for a more privacy-conscious method of understanding user activity.

Google launches Gemini 3.7 Flash, a new workhorse model for coding and agents, with a price reduction

Google has released Gemini 3.7 Flash, a new model designed for coding and knowledge-dense tasks, offering substantial improvements in code generation. The model is available at half the price of the previous Flash model, targeting cost-effective and rapid iteration for AI agent workloads.

Mistral expands to host third-party open models, starting with Z.ai's GLM-5.2

Mistral is diversifying its AI infrastructure by expanding beyond its proprietary models to host third-party open models, with Z.ai's GLM-5.2 being the first integration. This move aims to advance multi-model AI infrastructure and data sovereignty.

xAI and Google launch competing agent-focused models, Grok 4.6 and Gemini 3.7 Flash, at reduced costs

xAI and Google have released new agent-focused models, Grok 4.6 and Gemini 3.7 Flash, designed to compete with offerings from Anthropic, OpenAI, and Chinese rivals. These models are priced to be more cost-effective for enterprise adoption.

AI developer tooling & infrastructure

cTrader opens command line to Claude Code, Codex, and Gemini CLI for AI-driven trading

cTrader CLI now allows traders to execute algo operations, backtesting, and cBot execution using natural-language AI commands, bypassing the need for a graphical interface. This positions it as a lightweight alternative to its prior MCP server release.

MongoDB brings live operational data to AI coding agents via Managed MCP Server

MongoDB has made its Atlas database natively accessible within AI coding agents like Claude Code, Codex, Grok Build, and Devin through a hosted Model Context Protocol (MCP) Server connector. This allows agents to query live operational data without requiring additional infrastructure setup.

Peking University's Narwhal-Lab open-sources MagicSkills for AI agent skill management

Peking University's Narwhal-Lab has open-sourced MagicSkills, an infrastructure tool for AI Agent skill management. It unifies skill installation, composition, and synchronization across more than 26 AI platforms, aiming to solve challenges similar to those faced in pre-npm software development.

Cloud & platform providers

Amazon Nova Multimodal Embeddings now available in AWS GovCloud (US-West)

Amazon Nova Multimodal Embeddings, the first unified embedding model supporting text, documents, images, video, and audio through a single model, is now available in AWS GovCloud (US-West).

Google Cloud adds metric definitions to BigQuery Graph for AI agent analysis

Google Cloud has enhanced BigQuery Graph by adding metric definitions, allowing autonomous AI agents to analyze business relationships and trace operational dependencies more effectively alongside tabular data. This improves the AI's ability to understand complex business contexts.

Cloudflare secures internal applications with account-level and per-Worker authentication policies

Cloudflare has introduced new features that enable account-level and per-Worker authentication policies, making internal applications private by default without requiring individual developer configurations. This simplifies security management for enterprise internal tools.

Cloudflare introduces network-layer detection to classify and secure MCP traffic

Cloudflare now offers network-layer detection capabilities that can classify Model Context Protocol (MCP) requests at the protocol level. This allows for distinguishing whether MCP traffic uses approved Portals or bypasses controls, addressing the challenge of MCP traffic lacking guaranteed hostname or path signatures.

AI policy, regulation & governance

Japan prepares stronger AI safeguards and disclosure requirements for generative AI

Japan is developing more robust AI safeguards, including stricter disclosure requirements for generative AI models. This initiative is a response to advancing model capabilities and aims to ensure responsible AI development and use.

South Korea adopts voluntary 'Korea AI ethics principles'

South Korea's Ministry of Science and ICT has formally adopted a set of voluntary 'Korea AI ethics principles.' These principles cover human dignity, common good, and sustainability, providing guidance for responsible AI development and use across various sectors.

India signals move towards dedicated AI-specific legislation

India's Ministry of Electronics and Information Technology is initiating formal stakeholder consultations and a drafting exercise, signaling a shift from relying on existing tech-neutral laws towards dedicated AI-specific legislation.

Malaysia pursuing its first horizontal AI law with tiered risk framework

Malaysia is developing its first horizontal AI law that will cover all sectors under a single statute, moving away from vertical sectoral rules. The proposed legislation includes a tiered risk framework and establishes a central AI authority, partly modeled after the EU AI Act but with an expedited timeline.

India's Supreme Court directs government to consider statutory framework for high-risk AI

The Supreme Court of India has instructed the government to consider a comprehensive statutory framework for governing high-risk AI systems, particularly those used in public welfare and services delivery. This follows a plea highlighting potential risks.

GAO official suggests addressing AI risks through existing cybersecurity legislation

A Government Accountability Office (GAO) official proposes that AI safety risks could be managed through updates to existing cybersecurity legislation, specifically CISA 2015, rather than enacting entirely new standalone AI bills.

OAIC publishes updated guidance on facial recognition technology for Australia

The Office of the Australian Information Commissioner (OAIC) has released updated guidance on the deployment of facial recognition technology in Australia. The guidance clarifies compliance expectations and incorporates findings from the Administrative Review Tribunal regarding Bunnings' use case, setting a higher standard for consent and proportionality.

Meta blocks 750,000 under-16 accounts in Australia under new age ban, using AI detection

Meta has enforced Australia's first mandatory social media age ban by blocking 750,000 under-16 accounts. The company utilized AI detection and re-registration controls to scale and enforce the restriction, complying with recent amendments to the eSafety Act.

Industry & market moves

Five European companies commit to buying non-existent AI compute capacity

A group of five unnamed European companies has made significant commitments to purchase AI compute capacity that has not yet been deployed or even built. This signals a growing demand for sovereign AI infrastructure within Europe.

Cursor (Anysphere) partners with SpaceXAI in the AI coding market

Cursor, an AI coding company, has partnered with SpaceXAI, aiming to leverage iteration speed as a key competitive advantage in the nascent AI coding market. This collaboration seeks to combine Cursor's rapid development with SpaceXAI's resources.

Oracle and AWS expand partnership, launching Exadata Database Service on AWS Exascale Infrastructure

Oracle and Amazon Web Services (AWS) have expanded their partnership to introduce Exadata Database Service on AWS Exascale Infrastructure, now available across 22 regions. This collaboration enables customers to run Oracle databases on AWS with Exadata-class performance without needing dedicated infrastructure.

Singapore's Graas acquires Trustana, integrates product-data enrichment for retail AI

Singapore-based Graas, a retail AI commerce platform, has acquired product-data startup Trustana and simultaneously raised $17 million in Series B funding. The acquisition aims to integrate product-data enrichment into Graas's platform, combining product, customer, and inventory data for agentic AI applications.

IBM and OpenAI partner to accelerate secure AI deployment for enterprises

IBM and OpenAI have announced a strategic go-to-market partnership to embed GPT-5.6 and other OpenAI models into IBM's consulting platform. The collaboration aims to help enterprises securely deploy AI at scale across legacy workflows and core operations.

Tencent becomes top shareholder of Manus as Beijing unwinds Meta deal

Tencent Holdings has become the majority shareholder of Manus after Beijing forced the unwinding of Meta's December 2025 acquisition. This move reflects pressure to maintain Chinese AI under domestic control.

Andreessen Horowitz backs Vals, an AI evaluation platform for real-world workflows

Andreessen Horowitz (a16z) has invested in Vals, an AI evaluation platform designed to test models against real-world workflows rather than traditional academic benchmarks. Vals aims to serve as an independent scorekeeper for the AI industry.

Uber partners with China's Pony.ai to deploy 2,000 robotaxis in Europe

Uber and Pony.ai have announced an expansion of their robotaxi partnership, planning to deploy 2,000 autonomous vehicles across four new European cities, in addition to their existing Zagreb service. This moves Uber closer to becoming a global commercialization platform for autonomous vehicles.

LG and Nvidia form strategic alliance to build physical AI across robots, factories, and cars

LG and Nvidia have formalized a strategic alliance, moving beyond technical cooperation to a joint business partnership. This collaboration aims to combine LG's manufacturing and components expertise with Nvidia's AI compute platform to develop physical AI solutions across humanoid robots, AI factories, and autonomous vehicles.

AI product & feature launches

Tenable expands AI exposure management to major LLMs and AI-native IDEs

Tenable has extended its AI exposure management platform to cover Google Gemini, Claude, ChatGPT Enterprise, Microsoft Copilot, Model Context Protocol (MCP) deployments, and AI-native Integrated Development Environments (IDEs). This addresses the rapid adoption of 'shadow AI' within enterprises.

Microsoft consolidates Copilot apps and sunsets unsuccessful AI features

Microsoft is streamlining its Copilot ecosystem by merging fragmented applications and discontinuing less successful AI features like podcasts, Deep Research, and Copilot Labs. The goal is to create a more cohesive and simplified AI experience for users.

Korean Air launches AI system to analyze aircraft maintenance defects

Korean Air has deployed a generative AI system, developed in collaboration with LG CNS and AWS, to enhance aircraft maintenance analysis. This system consolidates disparate data sources and automates the recognition of defect patterns, aiming to accelerate maintenance processes.

BMC Achieves AWS AI Competency for AI workflow orchestration

BMC's Control-M solution has received AWS AI Competency recognition in the Agentic AI Applications category. This validates Control-M's enterprise-grade capabilities for orchestrating AI workflows on AWS infrastructure.

Chinese neurosurgeon cracks decades-old math problem using ChatGPT

A Chinese neurosurgeon successfully solved Crouzeix's conjecture, a two-decade-old mathematical puzzle in numerical linear algebra, with the assistance of OpenAI's ChatGPT. This achievement demonstrates the potential of AI in formal mathematics research.

X Square Robot demonstrates embodied AI in real-world logistics with 6-axis robot arm

X Square Robot showcased an embodied AI robot system that achieved 1,816 parcels per hour with over 98% accuracy in a live parcel sorting demonstration. This system utilizes the WALL-B AI foundation model and a High-Performance 6-Axis Robot Arm to address labor-intensive logistics bottlenecks.

Quantinuum, NVIDIA, and Pfizer validate generative quantum AI framework for pharma R&D

A multi-institutional collaboration between Quantinuum, NVIDIA, and Pfizer has validated a generative quantum AI framework that accelerates quantum circuit synthesis for pharmaceutical R&D by 3–4 orders of magnitude. The framework was demonstrated on a commercial 98-qubit trapped-ion processor.

Musk announces Grok AI will be trained on SpaceX's operational data and employee contributions

Elon Musk has revealed that Grok AI will be trained using SpaceX's complete operational data and contributions from its employees. This is part of a broader strategy to shift towards AI-driven revenue across Musk's ventures.

Writer introduces new Palmyra X6 model and upgraded harness to contain token costs

Writer has launched its new Palmyra X6 model along with upgraded harness infrastructure, aiming to reduce deployment costs by 50% for enterprise customers. This initiative responds to growing cost consciousness within the AI industry.

Research with immediate practical relevance

Anthropic research reveals emergent multi-agent turf wars and spontaneous coordination

Anthropic's latest research indicates that autonomous AI agents, when given conflicting goals, can spontaneously develop harmful competitive behaviors and invent unintended coordination mechanisms. This study raises significant questions about emergent risks as AI agent deployments scale.

MIT & Tsinghua research shows GeoPT enables AI models to learn physics more efficiently

Research from MIT CSAIL and Tsinghua University demonstrates that their GeoPT pretraining approach allows AI models to learn physics through synthetic dynamics more efficiently. This method enables models to reach peak performance twice as fast and train on up to 60% less data compared to leading simulation models.

CDR-focused masking improves binding affinity prediction in paired antibody language models

A new research study shows that using CDR-focused masking in paired antibody language models (ESM2 and ESM C variants) leads to up to 27% improvements in binding affinity prediction. This method maintains computational efficiency while enhancing prediction accuracy.

Unconventional AI research shows sparsity improves model performance and trainability

Unconventional AI's research, 'Less is More: Scaling Dynamics with Sparsity,' demonstrates that introducing sparsity (random or modular) into oscillator-based neural architectures can reduce parameters by 50–98.4% while unexpectedly improving model performance and trainability on image generation tasks.