The Informed Layer · Newsscan
Daily briefing

Model leaks and rollouts accelerate; regulators respond

← All issues

Anthropic / Claude ecosystem

Claude Users Report Brief Appearance of Mythos in Claude Code Despite Denial

Users claim that Anthropic's cybersecurity-focused Mythos model briefly surfaced in Claude Code and Claude Security, sparking speculation about a wider release, though Anthropic has denied any plans.

Anthropic Launches MCP Tunnels for Secure Internal Database Access by Claude

Anthropic introduced MCP Tunnels and Self-Hosted Sandboxes to allow Claude Managed Agents to securely access internal databases via outbound-only connections, enhancing enterprise security.

Frontier model providers

GPT-5.6 Leaks Hint at June Launch with Major Upgrades

Leaks suggest OpenAI's next large language model, GPT-5.6, could launch as early as June 2026, possibly with a Pro variant, sparking anticipation in the AI community.

Google Gemini 3 Ships 2.8x Faster and 47% Cheaper than GPT-5

Google has released Gemini 3, a new AI model that offers 2.8 times the speed and a 47% lower price compared to OpenAI's GPT-5, intensifying competition.

OpenAI Ads Manager Beta Adds Budgeting and Geo-Targeting

OpenAI's Ads Manager Beta now includes advanced budgeting, geo-targeting, aggregate reporting, and dynamic CTAs testing features.

Google Launches Gemini Code Assist and Code Architect to Rival Copilot

Google introduced upgraded AI coding tools, including the multi-file Code Architect, integrated with popular IDEs to challenge OpenAI and Anthropic in the coding assistant market.

Google Beam Introduces Life-Size AI Agents for Meeting Rooms

Google Beam is a new system that projects life-size AI agents into physical meeting rooms, enabling more natural AI-human collaboration.

OpenAI Releases Open-Weight Tool for Custom AI Safety Rules

OpenAI open-sourced 'gpt-oss-safeguard', a set of safety models that allow developers to tailor AI safety policies transparently for their applications.

AI developer tooling & infrastructure

OpenClaw Beta Release Cuts Latency and Tightens Delegation

OpenClaw version 2026.5.22-beta.1 reduces model-listing latency and improves deterministic delegation, enhancing the runtime efficiency of agent operations.

Cloud & platform providers

AWS MCP Server Reaches General Availability with IAM Governance

Amazon Web Services' MCP Server is now generally available, allowing AI agents to securely access AWS services with IAM-based governance and full API coverage.

AI policy, regulation & governance

Cohere Shapes Canadian AI Compute Funding Framework

Cohere played a pivotal role in designing a long-term AI compute funding framework with the Canadian government, anchoring a nationwide AI infrastructure strategy.

FTC Issues First Enforcement Fines Under TAKE IT DOWN Act for AI Misuse

The US Federal Trade Commission sent enforcement letters to eight online platforms, imposing fines for AI-generated deepfakes and non-consensual intimate imagery, marking the first major action under the Act.

Australia Raises Concerns Over Aged Care Algorithm Lacking Human Override

The Australian Coalition has sounded the alarm over an aged care assessment algorithm that operates without human override, raising questions about AI accountability in critical services.

Industry & market moves

Meta Mined Employee Keystrokes Before Cutting 8,000 Jobs

Investigation reveals Meta secretly deployed a keystroke-logging program to gauge AI replacement potential before laying off 8,000 employees, raising surveillance and labor concerns.

SpaceXAI Hiring Engineers and Physicists, No AI Experience Needed

Elon Musk announced that his combined SpaceX and xAI venture is hiring engineers and physicists without AI backgrounds, with Musk personally reviewing candidates.

Harbor Capital Files for Lab-Branded ETFs Targeting AI Giants

Harbor Capital plans to launch ETFs focused on specific AI labs like Anthropic, OpenAI, and xAI, offering public market exposure to private AI frontrunners.

Moment Raises $78M for AI Operating System in Wealth Management

Fintech startup Moment, founded by former Citadel quants, raised $78 million to build an AI agent platform for wealth management firms, automating client-facing processes.

LTIMindtree to Acquire Randstad Tech Operations for €160M, Expanding AI Capabilities

IT services firm LTIMindtree will acquire Randstad's technology business in Europe and Australia for up to €160 million, boosting its AI and defence portfolio.

AGAE Acquires HyalRoute for $2.31B in AI Optical Network Pivot

All In FutureTech Alliance (AGAE) is acquiring a controlling stake in Southeast Asian fiber optic company HyalRoute, pivoting to build an AI optical network platform combining optical computing and transmission.

AI product & feature launches

Chrome 149 Updates WebMCP Origin Trial with Token-Based Security

The WebMCP origin trial in Chrome 149 now requires origin-bound tokens, moving beyond the earlier Chrome 146 preview to enable secure website-to-AI agent interactions.

Genesis AI Unveils GENE-26.5 Model for Human-Like Robotic Dexterity

Genesis AI's new GENE-26.5 model aims to enable robots to perform complex, long-horizon tasks with human-like dexterity and adapt quickly to new environments.

Research with immediate practical relevance

Claude Sonnet 4.8 Leak Reveals Anthropic's Next Model

A leak in over 512,000 lines of code has revealed references to an unreleased Anthropic model, Claude Sonnet 4.8, hinting at upcoming capabilities and roadmap.

GPT-5.6 Shows UI Cleanup Breakthrough in Internal Build

OpenAI's unreleased GPT-5.6 reportedly generated a polished minimal-design app called Lumen Notes, indicating significant frontend-quality improvements.

Google DeepMind Selects EVE Online as Testbed for AI Agents

DeepMind partners with Fenris Creations to use the space MMO EVE Online as a continuous environment for training AI agents in long-horizon planning and continual learning.

Microsoft Releases Webwright, an Open-Source Terminal-Based Web Agent

Microsoft Research open-sourced Webwright, a web agent framework using Playwright scripts that achieves 60.1% on the Odysseys benchmark, outperforming a raw GPT-5.4 baseline of 33.5%.

KAIST Develops AI Interviewer for Psychiatric Intake in 30 Minutes

South Korean researchers created an LLM-based AI interviewer that collects comprehensive psychiatric intake data in half an hour, aiding clinicians.

Apple Open-Sources LiTo for Photorealistic 3D AI Content Creation

Apple made LiTo, a surface light field tokenization tool, publicly available, enabling developers to generate realistic 3D content from images with accurate lighting.