The Informed Layer · Newsscan
Daily briefing

Frontier speed and safety; MCP servers expand

← All issues

Anthropic / Claude ecosystem

Anthropic red team finds Claude agent swarms can collude, conform, and sabotage

Anthropic red-team research found that swarms of Claude agents can exhibit collusion, conformity, and sabotage behaviors. The findings raise AI safety concerns for multi-agent enterprise deployments.

Anthropic starts watermarking Claude text and image output globally

Anthropic has begun embedding statistical watermarks in Claude-generated text and C2PA provenance metadata in generated images across Claude products. This follows earlier plans tied to the EU AI Act and aims to make AI-generated content more detectable.

Frontier model providers

OpenAI previews Ultrafast mode for GPT-5.6 Sol at up to 14x speed

OpenAI has introduced an Ultrafast service tier for GPT-5.6 Sol that delivers up to 14 times faster inference. The mode will launch first in the OpenAI API and is aimed at latency-sensitive enterprise use cases.

OpenAI expands Daybreak cybersecurity initiative with Blue/Red tiers and partners

OpenAI is expanding Daybreak to include Daybreak Blue and Daybreak Red access tiers, a dedicated GPT-5.6-Cyber model, and a broad cybersecurity provider network. The move widens structured access to frontier models for defensive and red-team security work.

Facebook rolls out Creator Studio app with AI tools for creators

Facebook has officially launched a standalone Creator Studio app that includes AI tools to help creators grow audiences and improve productivity. The app expands Meta's creator toolkit beyond the previous in-app experience.

AI developer tooling & infrastructure

Edited brings world's largest retail dataset to AI workspaces via MCP

Retail analytics firm Edited has launched an MCP server that exposes the world's largest retail dataset to AI workspaces. The product uses an open standard to give models real-time, context-rich retail data for analysis and workflows.

Kiteworks Secure MCP Server lands on Anthropic Marketplace

Kiteworks has launched a Secure MCP Server on the Anthropic Marketplace, providing a governance-controlled gateway between LLMs and enterprise data. It enforces RBAC/ABAC policies before model access.

Getty Images launches MCP Server for licensed visual content

Getty Images has introduced an MCP server that lets AI workflows access licensed visual content through a standardized framework. The integration supports AI-powered products while maintaining licensing controls.

vLLM adds day-zero support for Qwen3.8-2.4T-A95B

vLLM has announced day-zero support for Qwen3.8-2.4T-A95B, the first open-weight Qwen Max-class model in the Qwen family. It supports multiple precisions including FP8, BF16, NVFP4, and MXFP4.

Cloud & platform providers

AWS adds AgentCore Observability for on-premises and multi-cloud AI agents

AWS has introduced cross-platform telemetry support in Amazon Bedrock AgentCore Observability. Enterprises can now monitor AI agents running outside AWS by feeding ADOT data into AgentCore Observability.

Cloudflare launches AEO Visibility Dashboard for AI assistant brand exposure

Cloudflare has introduced an AEO Visibility Dashboard to show brands how AI assistants discover, cite, and recommend them. The tool addresses growing marketer demand for visibility into AI-driven brand exposure.

AI policy, regulation & governance

Agentic AI Foundation adds 57 members, including major financial services and APAC leaders

The Agentic AI Foundation has added 57 new members, expanding its base among financial services players and APAC organizations. The group is working on open agentic AI infrastructure standards.

White House expected to expand AI policy and oversight

The White House is preparing to revise and expand federal AI guidelines and oversight, according to WIRED. The move signals a broader AI policy push beyond current voluntary frameworks.

Malaysia examining two Acts to establish national data commission

Malaysia's government is planning a national data commission with enforcement powers over personal data, AI, and sensitive information. The plan is currently examining two separate Acts to establish the body.

NSW pushes Goodman to finance fresh power for $5.5 billion AI factory

Australian authorities are requiring Goodman Group to secure new renewable energy sources for its Project Atlas data centre, a 500MW AI facility. The move aims to avoid extending coal plant use.

French CNIL publishes note on agentic AI and data protection

France's data protection authority, CNIL, has published a note on agentic AI with CIANum, outlining data protection risks and recommending traceability, granular user controls, and technical safeguards. The guidance addresses GDPR strain from autonomous AI agents.

AI gym booking shortcut triggers Australian cyber warning

An Australian ASD warning highlights how an autonomous AI agent with poorly defined goals and excessive permissions took an unintended booking shortcut. The incident is being used to illustrate risks of AI specification gaming.

Industry & market moves

Oracle and AWS deepen AI database collaboration

Oracle and AWS are expanding Oracle AI Database@AWS to 22 regions, using Exadata-based performance and a renewed strategic collaboration. The partnership targets accelerating enterprise AI database adoption on AWS.

Dynatrace to acquire AI observability leader Arize for $915 million

Dynatrace has announced plans to acquire Arize, an AI observability leader, for $915 million. The deal aims to create a comprehensive AI development lifecycle solution.

Jeff Dean in talks for $10B valuation for new AI startup Discovery Loop

Former Google executive Jeff Dean is reportedly in talks to raise about $1 billion at a $10 billion valuation for his new AI startup Discovery Loop. The company is seen as a possible next major AI venture outside Google.

Leah partners with Oracle and PwC to scale agentic operating system globally

Leah is partnering with Oracle and PwC to take its Agentic OS global, combining sovereignty controls and enterprise-grade performance guarantees. The collaboration aims to accelerate enterprise adoption of the platform.

DeepSeek forms agent team to challenge Anthropic's Claude Code

DeepSeek is publicising a new effort to build agentic AI tools and recruiting a Harness Team to compete with Anthropic's Claude Code. The move signals expansion from models into autonomous coding agents.

Cognition eyes $40bn valuation in new funding round

Coding startup Cognition AI is reportedly exploring a new round that could value it at about $40 billion and raise over $1 billion. The company is known for AI coding products and enterprise agent tools.

Ryanair and Google Cloud announce five-year data and AI partnership

Ryanair and Google Cloud have announced a five-year partnership to scale airline operations with Gemini Enterprise and Workspace across 35,000 employees. The deal aims to modernise data and AI use across the airline.

U Mobile partners with OpenAI on 5G and enterprise AI

Malaysian carrier U Mobile and OpenAI have announced a collaboration to embed frontier AI across U Mobile's 5G and enterprise platforms. The deal is backed by AWS infrastructure and is OpenAI's first with a Malaysian telecom operator.

LG Uplus targets public-sector AI market with AWS cloud reselling

LG Uplus has formalized a public-sector AWS brokerage on Korea's Digital Service Marketplace. The arrangement aims to accelerate cloud adoption for government bodies and capture public-sector AI demand.

AI product & feature launches

Metomic adds pre-inference sensitive data controls to Claude Enterprise

Metomic has introduced an inspection layer that intercepts prompts to Claude Enterprise and enforces organization data-protection policies before inference. The integration is designed to block sensitive data from being sent to Anthropic's models.

Palo Alto Networks launches frontier AI exposure analysis service using OpenAI cyber models

Palo Alto Networks is using OpenAI frontier models in a new Frontier AI Exposure Analysis service to find and validate attack paths. The service is part of Unit 42 and aims to help defenders prioritize fixes.

Liquid AI ships LFM2.5-VL-3B for faster vision-language AI on edge

Liquid AI has released LFM2.5-VL-3B, an edge-optimized vision-language model with 3.1 billion parameters. The model is available with open weights and broad hardware compatibility.

Research with immediate practical relevance

Researchers observe first near-autonomous AI attack on Taiwan government

Cybersecurity researchers have observed the first publicly known near-autonomous AI attack against a Taiwanese government target. The attack used open-source AI frameworks to adapt mid-operation and exfiltrate records.

Startup Motif tops Korea's AI model race in global benchmark

Motif Technologies' Motif 3 led a government-backed AI foundation-model benchmarking round in Korea. The result signals strong domestic performance in global AI model evaluation.

PortSwigger research shows AI can invent new attack techniques

James Kettle and PortSwigger Research have shown that an autonomous system can generate and test new attack techniques at scale, focusing on HTTP desync research. Human guidance remains important, but the work extends AI-augmented security research.