The Informed Layer · Newsscan
Daily briefing

Anthropic and Google drop models; MCP adoption spreads

← All issues

Anthropic / Claude ecosystem

Anthropic Launches Enterprise Frontier Safeguards for Zero-Data-Retention AI Monitoring

Anthropic has released Enterprise Frontier Safeguards (EFS), developed alongside more than 100 enterprise customers in heavily regulated sectors. The feature enables automated cross-session misuse monitoring while allowing organizations to retain customer-controlled encryption keys and store Claude logs in their own cloud infrastructure.

Anthropic Releases Claude Fable 5.1 and Cuts Cached Token Pricing by 75%

Anthropic has officially launched Claude Fable 5.1 and the restricted-access Claude Mythos 5.1. Fable 5.1 offers doubled agentic scientific capability, improved ambiguous task reasoning, and a 75% price cut on cached prompt tokens, reducing overall operating costs by up to 45% for sustained multi-agent workflows.

Anthropic Rolls Out Pre-Built Claude AI Blueprints for Retailers

Anthropic has launched pre-configured agent architectures designed for retail and ecommerce workflows ahead of the holiday shopping season. Built with partners like Shopify, Visa, and Mastercard, the shopper and merchant blueprints aim to automate complex catalog discovery and order handling.

Claude Fable 5.1 Solves 373-Year-Old Cryptographic Cipher in 44 Minutes

In a research capability demonstration, Claude Fable 5.1 cracked Urquhart's 1653 'Cyphral Distich' cipher in 44 minutes. The historical puzzle had resisted systematic cryptographic and computational analysis since 1899.

Frontier model providers

Google Releases Gemini 3.8 Flash and Specialized 3.8 Flash Cyber Model

Google has announced Gemini 3.8 Flash alongside Gemini 3.8 Flash Cyber, an offensive and defensive security variant designed for vulnerability detection and automated patch validation. The base 3.8 Flash achieves competitive coding benchmark scores on DeepSWE while preserving low-latency inference pricing.

Google Introduces Agentic Video in Gemini to Cut Processing Costs by 66%

Google DeepMind has introduced Agentic Video understanding across Gemini Flash models. The architecture dynamically inspects and transcribes video clips at variable frame rates based on conversational context, reducing token consumption by up to 88% and analysis costs by 66% while improving extraction accuracy by 7%.

AI developer tooling & infrastructure

Open-Source Go CLI 'codes' Enables File-Based Multi-Agent Orchestration for Claude Code

A new open-source Go CLI tool named 'codes' has been released, providing an MCP server with 43 tools to tame multi-agent Claude Code instances. The utility coordinates agent state and workflows using local YAML templates and file-based tracking without requiring dedicated database infrastructure.

Zoho Launches Catalyst 3.0 for Direct Production Deployment of AI Code

Zoho has launched Catalyst 3.0, a serverless cloud engine designed to push AI-written application code straight to production environments with integrated continuous testing, compliance scanning, and automated provisioning.

Accessibility Tracker Launches Model Context Protocol Connector for Claude

Accessibility Tracker has released an official Model Context Protocol (MCP) server integration, enabling developers to conduct WCAG conformance audits, triage remediation tickets, and update issue logs directly inside Claude chat.

Coder and SpaceXAI Partner on Self-Hosted Agentic Coding for Regulated Sectors

Enterprise workspace platform Coder has partnered with SpaceXAI to launch Coder Agent Relay for Cursor. The integration enables air-gapped and heavily regulated enterprises to run autonomous coding workflows within their private infrastructure without telemetry or source code exfiltration.

Yodeck Debuts MCP AI Agent for Natural-Language Digital Signage Control

Digital signage software provider Yodeck has launched an AI Agent supporting the Model Context Protocol. The integration allows operators to manage display networks, schedule media, and run diagnostics directly through ChatGPT, Claude, and Cursor.

Security Compass and ZeroPath Partner to Secure AI-Generated Code

Security Compass has integrated its SD Elements compliance platform with ZeroPath's autonomous security scanner. The combined solution automatically injects security requirements and conducts real-time verification as developers use generative coding assistants.

LlamaIndex and Kaggle Launch ExtractBench for Complex Document Extraction

LlamaIndex and Kaggle have released ExtractBench, an open benchmark evaluating AI extraction accuracy on 370 real-world enterprise documents across 4,869 pages. The benchmark evaluates model reliability on dense financial statements, scanned PDFs, and complex multi-page tables.

Open-Source Krypton MCP Proxy Introduces Zero-Trust LLM Defense and Merkle Auditing

Developers have released Krypton MCP, an open-source Go proxy that sits between LLM clients and backend tools. The framework implements real-time prompt injection filtering, dynamic PII masking, and cryptographic Merkle tree audit trails for AI agent executions.

Innovid Connects NIVO AI Platform to Meta Ads via Model Context Protocol

Advertising technology firm Innovid has integrated its NIVO AI system with Meta's Model Context Protocol interface, allowing enterprise marketing teams to orchestrate campaign modifications, query cross-channel metrics, and generate ad variations via natural conversation.

Knorex Unveils KAI Assist and XPO MCP Integration for Programmatic Advertising

Digital marketing provider Knorex has introduced KAI Assist and an MCP server for its XPO platform. The tool enables media buyers to connect models like Claude and ChatGPT directly to advertising bid infrastructure for automated campaign adjustments.

Cloud & platform providers

Google Reaches General Availability for BigQuery Graph to Ground AI Agents

Google Cloud has announced the GA release of BigQuery Graph, unifying relational tables and graph analytics within a single engine. The platform adds native GQL querying and vector indexing to enhance knowledge retrieval for enterprise AI agents.

AWS Introduces Amazon Quick for Natural-Language Enterprise App Generation

Amazon Web Services has launched Amazon Quick, a no-code service allowing business users to generate custom web applications from natural-language prompts. The tool connects securely to backends including Salesforce, Jira, and internal AWS data sources.

AI policy, regulation & governance

Apple has raised novel legal arguments regarding the difficulty of clawing back intellectual property and trade secrets once ingested into generative AI weights and autonomous agents. The filing emphasizes the technical limitations of current machine unlearning methods for enterprise IP protection.

FOIA Lawsuit Seeks Disclosure of Secret US Federal AI Safety Review Rules

Nonprofit Protect Democracy has filed a federal lawsuit against the Trump administration to force the public release of secret criteria and testing procedures used in government pre-release AI safety evaluations, including the GOLD EAGLE review framework.

G20 Summit Exposes Deep Divide Over US Pre-Deployment AI Testing Body

Discussions at the G20 AI summit revealed sharp disagreements over global AI oversight, with Google DeepMind CEO Demis Hassabis advocating for an independent US pre-deployment testing body while US administration representatives championed voluntary, decentralized industry standards under the 'Carolina Principles'.

UK Cyber Bill Explicitly Excludes AI Model Developers from Statutory Mandates

The UK Government has rejected proposed House of Lords amendments that would have imposed binding cybersecurity duties on AI foundation model developers under the Cyber Security and Resilience Bill, maintaining a voluntary governance model focused solely on downstream enterprise deployers.

Assistant Technology Minister Andrew Charlton stated that Australia's emerging AI copyright framework and data centre energy requirements are designed to balance economic growth without unnecessarily clashing with deregulatory shifts in the United States.

Australian Parliamentary Committees Warn Against AI-Generated Hallucinations in Inquiries

Chairs of Australian federal parliamentary committees have raised formal warnings after discovering multiple inquiry submissions containing fabricated citations and AI hallucinations, prompting calls for strict verification standards on public policy evidence.

The US Department of Justice has filed an amicus brief in the ongoing New York Times v. OpenAI copyright lawsuit, arguing that fair use protections should apply to ingesting copyrighted material for LLM training in order to protect national AI competitiveness.

Industry & market moves

Andreessen Horowitz Expands Growth Fund to $8.5 Billion to Scale AI Leaders

Venture firm a16z has expanded its Growth Fund (Fund V) to $8.5 billion. The expanded capital base will focus on late-stage enterprise AI foundation model developers, data platforms, and compute infrastructure providers.

Mistral AI Signs Strategic Sovereign AI Agreement with Côte d'Ivoire

Mistral AI has partnered with Côte d'Ivoire's Ministry of Digital Transition to conduct a comprehensive national AI audit and establish a sovereign deployment roadmap. The initiative will deploy localized open models across public administration, healthcare, education, and agricultural sectors.

Adobe Acquires Indian Marketing Intelligence Startup Rilo to Expand CX Suite

Adobe has acquired Rilo, an Indian marketing intelligence startup, for an undisclosed sum. Rilo's automated analytics and campaign intelligence agents will be natively integrated across Adobe Experience Cloud.

Physical AI Startup Lyte Raises $165M Series C at $1.6B Valuation

Lyte, a physical AI robotics perception startup founded by former Apple engineers, has raised $165 million in Series C funding led by Maverick, valuing the company at $1.6 billion. The capital will support large-scale deployment of its spatial foundation models for industrial robotics.

HiddenLayer Raises $100M Series B for AI Agent Runtime Defense

AI security provider HiddenLayer has secured $100 million in Series B financing to accelerate the development of its runtime protection platform, expanding its defensive toolset against prompt injection and unauthorized execution in autonomous coding agents.

AfterQuery Reaches $3.2B Valuation in Series B Funding Round

Enterprise data platform AfterQuery has completed a Series B round valuing the company at $3.2 billion, becoming Y Combinator's fastest-ever startup to reach unicorn status just 18 months from its founding.

Conveo Raises $50M Series A Led by DST Global and Balderton for AI Consumer Research

AI consumer insights startup Conveo has raised $50 million in a Series A round co-led by DST Global and Balderton Capital. The platform uses conversational video AI agents to conduct, transcribe, and analyze qualitative consumer interviews at scale.

AI product & feature launches

Forescout Demonstrates Claude Porting Industrial PLC Exploits for $536

Security researchers at Forescout's Vedere Labs showed that Claude could be guided to port an authenticated remote code execution exploit (CVE-2021-31886) between two different WAGO PLC hardware models. The automated porting process required $536 in API tokens and 8 hours of execution, though an unassisted attempt to build a custom C2 implant resulted in bricking the test device.

B.AI Transitions DeepSeek V4 Flash to Paid Tier with 50% Baseline Pricing

Singapore-based API platform B.AI has concluded free trial access for DeepSeek V4 Flash and V4-Flash-Vision-Exp. The platform has instituted a 50% discounted baseline billing rate alongside deep discounts for off-peak workloads to manage server load.

Sonos Launches Hardware Lineup with Open Model Context Protocol Integration

Sonos has unveiled its Sonos 27 operating system alongside the Beam Ultra and Ace Ultra headphones, featuring an open MCP server endpoint. The integration allows third-party AI agents from OpenAI, Google, and xAI to control audio hardware without relying on closed proprietary assistants.

Owkin Licenses K Pro AI Scientist Platform to Boehringer Ingelheim

Biotech AI firm Owkin has signed an expanded licensing agreement with Boehringer Ingelheim for its K Pro AI Scientist platform. The multi-year deal grants Boehringer Ingelheim access to multimodal oncology and immunology datasets alongside generative biological discovery models.

World Labs Unveils Atlas Multimodal 3D World Foundation Model

Fei-Fei Li's startup World Labs has announced Atlas, a unified spatial foundation model natively trained on text, images, video, and 3D data. Atlas generates interactive 3D simulations and photorealistic environments from single 2D images with precise, instruction-guided camera controls.

CrowdStrike Introduces SafeMind Closed-Loop Autonomous Red/Blue AI Defense

CrowdStrike has launched SafeMind, a dual-agent closed-loop security system that deploys paired red team (Red Tempest) and blue team (Blue Solano) agents against digital twins of enterprise infrastructure to autonomously identify and patch vulnerabilities before exploitation.

Research with immediate practical relevance

Puffin-World Model Released with Native Physics-Grounded 3D World States

Researchers have released Puffin-World, an open unified multimodal world model that grounds 3D spatial generation in physical geometry rather than pixel extrapolation, enabling stable long-trajectory camera simulation and physics-consistent environment rendering.