The Informed Layer · Newsscan
Daily briefing

AI agents go to work; governments watch

← All issues

Anthropic / Claude ecosystem

Anthropic launches Inference Hooks to let enterprises block sensitive data in Claude prompts

Anthropic introduced Inference Hooks, a new enterprise security layer that inspects all prompts and tool calls via a customer-operated DLP server before they reach Claude. Available across all Claude enterprise surfaces, it allows real-time scanning and blocking of sensitive data.

Claude Code v2.1.222 fixes session isolation and proxy connectivity

A minor update to Claude Code resolves session isolation issues, improves proxy connectivity, and fixes usage attribution. The release follows v2.1.221 and addresses specific bugs reported by users.

Frontier model providers

OpenAI adds education plugins to ChatGPT Work and Codex

OpenAI launched three education-focused plugins for K-12 teachers, higher education faculty, and college students that integrate with ChatGPT Work and Codex. The plugins aim to streamline lesson planning, grading, and research workflows.

OpenAI's GPT-5.6 models go live for U.S. government use via FedRAMP

OpenAI made its GPT-5.6 family (Sol, Terra, Luna) available to U.S. federal agencies through ChatGPT Enterprise, which is FedRAMP authorized. This marks the first time advanced OpenAI models are officially offered for government workloads.

OpenAI reduces GPT-5.6 pricing and adds Sol Fast mode

OpenAI announced lower prices for GPT-5.6 Luna and Terra models, alongside a new Sol Fast mode that offers faster inference at reduced cost. The move aims to improve cost efficiency for developers.

Meta makes AI Developer Assistant the primary support channel for developers

Meta transitioned its developer support to primarily use its AI Developer Assistant, automating responses and guidance for integrating with Meta's platforms. The move streamlines support and encourages developer AI adoption.

Cohere targets 1 trillion-parameter model with enterprise sovereign AI push

Cohere plans to launch Command A Plus, a 1-trillion-parameter model, by year-end as part of its enterprise sovereign AI strategy. The model is optimized for private, on-premise deployments for regulated industries.

DeepSeek V4-Flash tops global API usage, claims much lower costs

DeepSeek's V4-Flash model has become the most used API globally, with the company claiming significantly lower inference costs compared to competitors. If sustained, it could shift developer preferences toward lower-cost models.

AI developer tooling & infrastructure

VS Code 1.132 sharpens AI agent workflows with side chats and browser comments

The latest Visual Studio Code update introduces side chats and element-level comments in the browser, enhancing agent workflows. These features improve collaboration between developers and AI coding assistants.

Uber's ADR AI agent security project goes open source

Uber open-sourced its ADR (AI Agent Security) stack, which provides runtime security for autonomous AI agents. The project aims to become the industry standard for agentic security.

RouterBase unified AI API now connects 200+ frontier models

RouterBase launched a unified API compatible with OpenAI's format, giving developers access to over 200 frontier models with automatic fallback and consolidated billing.

Cloud & platform providers

AWS open-sources Kiro Crew to orchestrate autonomous AI engineering teams

AWS released Kiro Crew, an open-source framework that enables multiple AI coding agents to work together as an autonomous engineering team. It supports self-hosted deployments for governance and compliance.

Cloudflare launches WriteGuard for fine-grained MCP server controls

Cloudflare introduced WriteGuard, a tool that adds fine-grained access controls to MCP servers, including request-context-aware behavior such as pass-through, scrubbed audit, and risk-based blocking.

AWS unveils Prism prototype for distributed generative AI across government data

AWS shared a prototype called Prism that demonstrates six architectural patterns for deploying generative AI across distributed datasets without central consolidation, using fine-grained access controls via TOLAP.

Cloudflare OS launches as first AI workspace built around how companies work

Cloudflare OS is an open-source, zero-trust AI workspace that organizations can deploy on-premises, connecting internal systems and customizing the platform. It aims to integrate AI agents into existing workflows.

AI policy, regulation & governance

AI leaders propose SAFE guidelines for confidential cybersecurity incident sharing

The Open Secure AI Alliance introduced SAFE (Shared AI Findings Exchange) guidelines to enable confidential collection and analysis of AI incidents, aiming to reduce systemic cybersecurity risks in agentic systems.

Britain signals openness to AI regulation if voluntary safeguards fall short

A UK minister stated that the government would consider mandatory AI regulation if industry-led voluntary measures prove inadequate. The statement marks a shift from the previous purely pro-innovation stance.

Prime Minister Anthony Albanese announced plans to strengthen copyright laws to ensure creators are licensed and compensated when their work is used to train AI models. The move responds to growing concerns from the arts community.

Industry & market moves

Google DeepMind reshuffles top leadership; Demis Hassabis steps down as CEO

Google DeepMind CEO Demis Hassabis will move to Alphabet Chief Scientist, while Jeff Dean leaves to start a new AI company. The shakeup signals a strategic refocus on AGI and governance changes within Google's AI division.

DeepSeek reportedly reopens talks on RMB50 billion second funding round

Chinese AI lab DeepSeek is in discussions for a second funding round, targeting about RMB50 billion with a pre-money valuation of RMB500 billion. The funds will support model R&D and compute infrastructure.

AlixPartners acquires agentic AI consulting firm Artium

Global consulting firm AlixPartners acquired Artium, a specialist in agentic AI software consulting, to bolster its capabilities in delivering enterprise AI solutions and execution.

Google in talks for $1.5 billion-plus deal with AI coding firm Mechanize

Google is negotiating a deal valued at over $1.5 billion to acquire talent and license technology from Mechanize, a firm specializing in AI coding and model evaluation.

Legal technology provider BigHand acquired Ayora, an AI-driven pricing platform, to enhance its capabilities in data-driven legal pricing strategies using machine learning.

QuantHealth raises $45M to simulate clinical trials with AI

QuantHealth secured $45 million in funding to expand its AI platform that simulates clinical trials before patient enrollment, aiming to reduce costs and timelines in drug development.

AI product & feature launches

Former DeepMind team launches Reimagine Robotics for on-the-job learning robots

A team of ex-DeepMind researchers founded Reimagine Robotics, an AI robotics startup creating robots that learn tasks directly from worker demonstrations. The approach simplifies training for changing factory environments.

Tenable expands AI exposure management to cover Google Gemini, MCP, and AI IDEs

Tenable One AI Exposure now provides visibility and risk assessment for Google Gemini, MCP deployments, and AI-native IDEs. The platform aims to unify AI security posture across cloud, on-premise, and developer tools.

Seamless.AI partners with ChatGPT to integrate verified B2B sales data

Seamless.AI launched an integration that embeds its verified B2B contact and company data directly into ChatGPT, enabling users to perform prospecting and lead generation within the chatbot.

Lockheed Martin Skunk Works AI autonomously flies 27 live intercepts on X-62 VISTA

A Lockheed Martin Skunk Works AI agent successfully conducted 27 autonomous fighter intercepts on the X-62 VISTA test aircraft, using live sensor data and closing the kill chain in real time.

Hark previews Handoff, a fast and affordable browser-use AI agent

AI startup Hark unveiled Handoff, a computer-use agent that can autonomously navigate websites and complete tasks, claiming faster and cheaper performance than competitors like Adept or OpenAI's Operator.

Research with immediate practical relevance

XunZi AI biologist reveals disease-modifying therapeutic targets in Nature paper

A research article in Nature Biomedical Engineering introduces XunZi, an AI system that autonomously identifies disease-modifying targets by integrating multimodal data. It validated targets like CHK2 and IRAK4 in Parkinson's models.