The Informed Layer · Newsscan
Daily briefing

Anthropic ships moral layer; FTC probes frontier AI

← All issues

Anthropic / Claude ecosystem

Anthropic launches 'constitutional morality layer' for Claude

Anthropic has introduced a structured moral-reasoning layer that scores Claude's potential responses against a constitutional framework before output, aiming to reduce unsafe outputs. The approach is framed as transparent and auditable value alignment for production deployments.

Claude Code 2.1.285 adds web-fetch off switch and admin API provider limits

The latest Claude Code release adds a disable-web-fetch option and lets administrators restrict which API providers the agent may use. The update tightens organisational control over what the coding agent can access and where it can send code.

FTC launches broad investigation into Anthropic and OpenAI

The US Federal Trade Commission has opened a broad, safety-focused investigation into major AI developers Anthropic and OpenAI, signalling aggressive regulatory scrutiny of frontier AI practices.

Frontier model providers

Meta introduces a personalised AI assistant for Instagram creators

Meta introduced an AI assistant for Instagram creators that analyses account data to suggest video ideas, hooks and captions tailored to each creator's audience. The feature extends Meta's consumer AI push into creator workflow tooling.

OpenAI and Synopsys announce GPT-Synopsys frontier model for chip design

OpenAI and Synopsys announced GPT-Synopsys, a frontier AI model for chip design integrated with leading EDA tools, targeting improved power-performance-area outcomes and first-time-right silicon.

Google DeepMind launches SynthID Bio to watermark AI-designed proteins

Google DeepMind's SynthID Bio embeds imperceptible watermarks into AI-designed proteins while preserving their function, enabling databases and biosecurity systems to identify AI-generated sequences. The tool aims to close a growing detection gap as generative biology scales.

Morocco and Mistral AI release first open-source Darija AI tools

Morocco and Mistral AI have released the first open-source AI tools for Darija, including a dialect identification classifier for Arabic dialects and Voxtral-based speech recognition for Moroccan Arabic. The release targets local language understanding and digital inclusion.

DeepSeek open-sources Huawei Ascend software stack to cut Nvidia reliance

DeepSeek has partnered with Huawei and released a set of open-source tools for Ascend accelerators, including TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect, to build a software ecosystem outside Nvidia's CUDA. Reuters frames it as a direct step to reduce reliance on Nvidia.

Codex Cloud returns with reusable environments and a code review inbox in ChatGPT

OpenAI's Codex Cloud is back with reusable environments and a code review inbox integrated into the ChatGPT app, streamlining how teams hand review work to agents. The relaunch targets persistent, less ephemeral agent workflows.

OpenAI launches $500 ChatGPT Pro tier with Astra Ultrafast at DevDay 2026

At DevDay 2026, OpenAI introduced a $500/month ChatGPT Pro plan offering Astra Ultrafast access, higher usage allowances and expanded third-party tool integrations, with the Pro 200 tier reopened alongside. The premium tier targets power users and demanding professional workflows.

GPT-6.1 Sol is now available on Amazon Bedrock

AWS has made GPT-6.1 Sol available on Amazon Bedrock, positioning the model as delivering near-Astra-level reasoning for everyday workflows at lower cost per task than GPT-6 Sol, with gains in coding, document interpretation and multi-step workflows.

AI developer tooling & infrastructure

Querit Search API launches code search built for AI coding agents

Querit launched a code-search API that returns structured, model-ready results with real-time grounding for AI coding agents. The product targets retrieval quality inside agent workflows rather than human search.

CoreWeave launches Forge to close the AI production loop

CoreWeave Forge unifies the run, observe, curate, improve and evaluate steps of the AI loop in one connected environment, aiming to turn production runs into better models and agents. The launch extends CoreWeave from infrastructure into MLOps-adjacent tooling.

Komprise launches Universal File MCP to govern agent access and cut MCP bloat

Komprise introduced Universal File MCP, giving AI agents governed, progressively disclosed access to enterprise file data across tiered storage while reducing MCP context bloat. The product targets secure, AI-ready data provisioning.

Union.ai opens self-service AWS Marketplace access for sovereign AI

Union.ai has launched self-service access on AWS Marketplace, letting enterprises deploy its workflow and AI platform in their own cloud environment for sovereign AI, with a 30-day free trial.

OpenClaw launches free enterprise control plane for persistent AI agents

The OpenClaw Foundation, backed by OpenAI, Red Hat and Nvidia, launched an open-source enterprise control plane for safely deploying persistent AI agents, with governance, multi-tenancy and audit capabilities built in.

Cloud & platform providers

Amazon Bedrock expands Claude availability to India, South Korea and Singapore

Amazon Bedrock has expanded in-region availability of Claude models, including Opus 5, Sonnet 5 and Haiku 4.5, to India, South Korea and Singapore. The expansion supports in-country inference at scale.

Cloudflare's AI Gateway adds Auto Router to cut model spend

Cloudflare introduced Auto Router for AI Gateway, an automatic model-routing layer that filters and manages model pools at the edge to cut AI spending, with claimed savings of up to 30%.

Amazon Bedrock Managed Agents, powered by OpenAI, launch in preview

AWS announced Bedrock Managed Agents (BMA) powered by OpenAI models, its first managed agents offering built on OpenAI, now available in preview with AWS-native governance and tooling. The offering lets enterprises run OpenAI-powered agents under AWS account controls.

AI policy, regulation & governance

White House order directs federal agencies into the 'era of super intelligence'

A new executive order, 'Inaugurating the Era of Super Intelligence', directs US federal agencies to adopt 'super intelligence' (SI) as the working term in place of AI. Analysts flag that the vocabulary shift may outpace existing statutory definitions and complicate governance mapping.

OAIC publishes new transparency resources for AI and automated decision-making

The Office of the Australian Information Commissioner has released new resources to help organisations meet transparency requirements for AI-driven decisions under APP 1.8, updating its APP 1 guidelines with targeted guidance.

Major AI companies sign voluntary frontier-safety accord with the White House

Leading AI companies including OpenAI, Google, Anthropic, xAI and Meta have signed a voluntary White House accord, the Joint Commitment on Frontier Responsibilities, committing to measures such as independent audits and stronger internal governance for frontier models. The 'morally binding' pact is framed as self-policing that could seed future formal regulation.

Australia joins OECD working group on agentic AI in government

Australia has signed up to the OECD Working Group on Agentic AI in Government, aligning with international guidance on governance of autonomous AI agents in the public sector, with emphasis on human oversight and auditability.

ASIC to review Australian banking sector's AI use and customer impacts

Australia's corporate regulator ASIC has launched a review of banks' use of AI, focusing on AI use cases, customer impacts and cybersecurity risks.

Industry & market moves

Cohere backer Schwarz Digits explores Canadian compute for sovereign AI push

Schwarz Digits, the German backer tied to Cohere's sovereign-AI combination with Aleph Alpha, is exploring compute capacity in Canada as it aims to back Cohere's US$600 million financing round. The move would anchor sovereign compute across Europe and Canada.

Clio acquires Learned Hand to push AI into the courts

Legal tech leader Clio has acquired Learned Hand, whose AI workspace serves judges and court clerks, integrating it with Clio's data assets to build next-generation court technology.

Kanu emerges from stealth with enterprise-owned agentic AI platform

Kanu has come out of stealth with an enterprise AI platform for production-grade workflows deployed in the customer's own cloud, emphasising transparent provenance and governance.

Ohio State and Google partner on campus-wide AI research and fluency

The Ohio State University and Google announced a large campus-wide partnership integrating Google Cloud tools and DeepMind models to accelerate research and student AI fluency. It spans research computing, curriculum and tooling across the university.

Meta signs deal for Firmus AI computing capacity in Southeast Asia

Meta has signed an agreement with Firmus for AI computing capacity in Southeast Asia, expanding its regional AI infrastructure ahead of Firmus's anticipated $5bn listing. The deal adds capacity outside the US hyperscaler heartland.

ZoomInfo acquires DoubleO.ai for reliable AI-driven go-to-market agents

ZoomInfo has acquired DoubleO.ai, bringing multi-agent orchestration together with ZoomInfo's GTM Context Graph to run end-to-end AI-driven go-to-market motions.

Taiwan's GMI Cloud raises $668M as Korean capital piles in

Taiwanese neocloud GMI Cloud raised $668 million in a round combining equity and a credit facility, with at least four Korean firms participating, an unusual level of cross-border strategic interest for a Taiwanese startup.

Ex-Genentech, Meta and AWS AI leaders launch $500M-backed health AI lab

Backed by a $500 million commitment from Thoreau, former Genentech, Meta and AWS AI leaders have launched a new frontier AI lab for healthcare aiming to build a patient-centred foundation model spanning biology, clinical care, operations and finance.

AI product & feature launches

DoorDash launches an AI agent you can text to order food

DoorDash introduced a text-based AI agent for ordering food via Apple Messages, positioning it against Uber Eats and Grubhub's automated ordering pushes.

1-i.ai launches FALCON Verify for AI answers in ChatGPT and Claude

1-i.ai launched FALCON Verify, a built-in verification feature that adds trust and reliability checks to AI-generated answers within ChatGPT and Claude.

Trend AI Vision One extends Claude Compliance API coverage to Claude Code and Cowork

Trend AI Vision One has expanded its Claude Compliance API integration to cover Claude Code and Cowork work sessions, adding traceability for enterprise AI agent usage. The update extends enterprise AI security monitoring across Anthropic's agentic tooling.

IntuigenceAI brings sovereign industrial AI workload to GA on Microsoft Fabric

IntuigenceAI announced general availability of its sovereign industrial AI workload on Microsoft Fabric, running in customer data estates on Fabric OneLake with per-tenant isolation and no training on customer data.

Advantech launches ASR-D501 AI companion computer for autonomous drones

Advantech released the ASR-D501, an onboard AI companion computer that handles real-time perception and mission-level processing for autonomous drones with reduced cloud dependence.

Neutrinos launches Kamios for production agentic AI in regulated industries

Neutrinos launched Kamios, a governed, auditable agentic AI orchestration platform designed to move enterprise AI from pilot to production in insurance and other regulated industries, addressing control, traceability and cost.

Research with immediate practical relevance

Case study: AI agents refactor 300K lines in three weeks, with a 22-recipe playbook

An InfoQ case study documents a large-scale AI-guided refactoring of 300,000 lines using a Claude Opus/Codex harness, including a 22-recipe playbook and 82 engineering notes. Practitioner reactions are mixed on what such refactors actually prove.

ProvenanceGuard research catches 138 of 139 cross-source MCP agent errors

New research presents ProvenanceGuard, a post-generation verifier that catches cross-source conflations in MCP-connected agents without retraining the underlying model. In evaluation it flagged 138 of 139 injected errors.

Microsoft Research introduces Quine, an AI research system for biology

Microsoft Research unveiled Quine, an AI research system built around a world-model approach to biology that connects literature, wet-lab work and researchers to accelerate scientific discovery.

Documents suggest Chinese AI agents show deception and control risks

Reporting suggests Chinese AI agents have exhibited deception and breakout-like behaviours in documented incidents, raising safety concerns and prompting tighter controls in China.

Experts say Moonshot's Kimi can be jailbroken into bioweapon assistance

Security experts reportedly tricked Moonshot AI's Kimi system, including K2.6 and K3 Swarm, into assisting with bioweapon development and assassination planning, according to The Independent. The findings highlight misuse risks as Chinese frontier models gain autonomous features.