The Informed Layer · Newsscan
Daily briefing

AI agents and funding surge; policy disclosures mount

← All issues

Anthropic / Claude ecosystem

Payward joins Anthropic's Project Glasswing AI cyber defence program

Kraken parent Payward has joined Anthropic's restricted Project Glasswing to use Claude Mythos 5 for proactive vulnerability discovery and defence in financial infrastructure.

Frontier model providers

OpenAI institutes new safeguards after Hugging Face breach

OpenAI implemented safeguards for development/testing security monitoring and post-training alignment after a Hugging Face breach.

Mistral ESR experiment will host data in French compute capacities

The French higher education Mistral ESR experiment will keep data hosted within French compute capacities, according to Amue.

DeepSeek launches peak/off-peak pricing; Alibaba releases AI music model

DeepSeek introduced new peak/off-peak pricing for V4 models, and Alibaba released HappyShrimp 1.0, an AI music model.

DeepSeek Harness enters internal beta with open-source developer recruitment

DeepSeek Harness started internal beta and opened recruitment to open-source developers.

AI developer tooling & infrastructure

Cursor launches Origin code hosting service to compete with GitHub

Cursor launched Origin, a cloud code-hosting service with Git repository management, desktop and CLI access, and AI-assisted workflows designed to compete with GitHub.

LangChain introduces AgentCore Payments middleware

LangChain released AgentCore Payments middleware, enabling payment workflows within LangChain agents.

Light Cloud lets developers ship apps without leaving Claude Code

Light Cloud introduced an MCP server integration that enables full-stack deployment, live URLs, and automated build/ship/redeploy cycles directly inside Claude Code.

NEAR Protocol IronClaw 1.2 introduces shared team AI agents

NEAR AI IronClaw 1.2.0 introduces shared-access, login-free AI agents for teams in Slack and Telegram.

Google introduces formal verification for CEL

Google released a formal verification framework for CEL, enabling mathematical proofs of correctness for CEL expressions and policies in agentic systems.

xpander wants enterprises to own their AI agent control and context layer

xpander offers a vendor-neutral control plane and Universal Harness to govern enterprise AI agents with identity, memory, and runtime infrastructure.

Cloud & platform providers

Megazone Cloud and AWS jointly release three enterprise AI agents

Megazone Cloud and AWS released three enterprise AI agent solutions—AAG, ADT, and SPACE LENS—now listed on AWS Marketplace for global use.

AWS DynamoDB added native vector search, enabling embeddings and approximate nearest-neighbor queries directly within DynamoDB.

Cloudflare enhances security and visibility for MCP traffic

Cloudflare added governance and visibility components for AI agent MCP traffic in Zero Trust Gateway, including identification headers, dashboards, and Traffic Source selectors.

Cloudways launches managed infrastructure for OpenClaw and Hermes AI agents

Cloudways announced general availability of managed infrastructure for the open-source OpenClaw and Hermes AI agents.

AI policy, regulation & governance

French government to hire French AI providers such as Mistral

France signalled it will work with sovereign AI providers like Mistral for government procurement, excluding OpenAI amid recent cyberattack disclosures.

Japan approves draft Principle Code urging AI developers to disclose training data

Japan's AI expert panel approved a draft Principle Code urging voluntary public disclosure of AI training data without binding penalties.

FDA opens GenAI medical device feedback docket to October 19

FDA opened a public feedback docket on risk assessment, premarket evaluation, and postmarket monitoring of GenAI-enabled medical devices.

China tightens border crossings after Meta's play for AI startup Manus

China is introducing entry/exit controls effective September 15 to prevent leaks of data and advanced technologies, following the Meta-Manus episode.

Poland becomes first EU member state to bring AI Act implementing law into force

Poland brought its AI Act implementing law into force, though its national regulator is not yet established, according to a tracker.

UK ICO gains statutory duty to produce AI code of practice

Under new regulations, the UK ICO now has a statutory duty to produce a Code of Practice on Artificial Intelligence and Automated Decision-Making, affecting UK businesses.

EU AI Office opens three complaint routes covering Google and Meta systems

The EU AI Office opened three formal complaint routes under the AI Act, including a general tool, a downstream provider channel, and an anonymous whistleblower tool.

EU AI Act deadline for medical devices moves to August 2028

The EU AI Act high-risk compliance deadline for medical devices was moved to August 2028 under Regulation (EU) 2026/1744.

Industry & market moves

Meta taps disability groups to explore uses for AI glasses

Meta is allocating nearly $2 million across 30 organizations in 18 states to explore accessibility and independence uses of Meta AI glasses.

Oracle AI Database@AWS expands to 22 regions

Oracle expanded Exadata Database Service on AWS to 22 regions, increasing data migration and cloud tooling options for customers.

Wispr raises $280M at $2B valuation to build voice layer beneath every app

Wispr raised a $280 million Series B at a $2 billion valuation from Menlo Ventures to embed voice as a foundational layer across applications.

Stigg acquires Received.ai to add contract management and invoicing

Stigg acquired Received.ai to integrate contract management and invoicing into its runtime AI usage platform.

FORT Robotics to go public via SPAC to advance physical AI safety

FORT Robotics agreed to a business combination with Newbury Street II Acquisition Corp, valuing the combined entity at about $556.6 million post-money.

Groq raises $350M to fuel pivot from AI chips to neocloud

Groq raised $350 million at a $3.5 billion valuation to expand from AI chips to neocloud infrastructure.

AI product & feature launches

SiftedAI now available in the Claude connectors directory

Sifted listed SiftedAI for Claude in the Claude connectors directory, allowing Claude users to access SiftedAI directly.

GenBio launches a virtual cell AI model

GenBio AI launched AIDO Cell, a virtual cell world-model that simulates natural states and responses to perturbations across biological hierarchies.

Razorpay to launch Vulcan, India's first AI payments foundation model

Razorpay announced Vulcan, India's first AI payments foundation model, built with NVIDIA and AWS and trained on trillions of payment data points.

PolyAI launches direct integration with Epic

PolyAI launched a direct integration with Epic EHR, with PDS Health among the first to deploy at scale.

Google and UK Government launch AI contrail avoidance trial

Operation Blue Skies is testing AI-guided flight-path adjustments across Shanwick airspace to reduce aviation's non-CO2 climate impact from contrails.

Qwen3.8-27B runs frontier-class coding agents locally

Alibaba released Qwen3.8-27B, an Apache-2.0 open model that runs frontier-class coding agents and reasoning locally without a cloud API.

Alibaba launches HappyShrimp 1.0 AI music model

Alibaba publicly released HappyShrimp 1.0 as a consumer-facing AI music model.

Research with immediate practical relevance

DiG-bench and Faraday mark slow progress toward recursive self-improvement

New benchmarks DiG-bench and Faraday indicate frontier models still struggle on the hardest discovery and replication tasks, though some show gains.

MIT researchers find AI models get harder to attribute as they grow

MIT researchers found that generative diffusion model outputs become less attributable to training data as model size increases.

ByteDance Seed and Tsinghua AIR introduce CUDA Agent

ByteDance Seed and Tsinghua AIR introduced CUDA Agent, a large-scale agentic RL system that generates CUDA kernels faster than torch.compile.

Tencent releases open-weight GUI agent UI-Mate-27B

Tencent released UI-Mate-27B, an open-weight GUI agent with 77.0 OSWorld and live-screenshot reasoning, though no public API is available yet.