The Informed Layer · Newsscan
Daily briefing

Anthropic agents advance; AI security flaws exposed

← All issues

Anthropic / Claude ecosystem

Anthropic Releases Claude Opus 4.7 with Enhanced Agentic Capabilities and Improved Vision

Anthropic officially launched Claude Opus 4.7, focusing on improved agentic workflows with longer task handling, better instruction following, and self-verification. The update also features over three times higher vision resolution (up to 3.75 megapixels), a new 'xhigh' effort level for API users, and 'ultrareview' for Claude Code sessions to flag human-detectable issues. Anthropic also revealed its more powerful Mythos model scores significantly higher on benchmarks but is being withheld due to cybersecurity concerns.

Anthropic Expands London Operations, Plans to Offer Mythos AI to UK Banks Next Week

Anthropic is undertaking a major expansion in London, securing new office space for 800 people. Concurrently, the company plans to make its powerful, cyber-vulnerability-finding Mythos AI model available to UK financial institutions within the next week as part of its 'Project Glasswing' initiative.

Frontier model providers

Google DeepMind Unveils Gemini 3.1 Flash Text-to-Speech Model

Google DeepMind has released Gemini 3.1 Flash TTS, a new text-to-speech model offering enhanced controllability, expressiveness, and output quality. It's available in preview for developers via the Gemini API and Google AI Studio, and for enterprises through Vertex AI. The model supports over 70 languages, features multi-speaker dialogue, and includes audio tags for granular control over pacing, tone, and vocal style. All output is embedded with SynthID watermarking for authenticity.

Microsoft Releases MAI-Image-2-Efficient Model, Accelerating Shift Away from OpenAI

Microsoft has launched MAI-Image-2-Efficient, a faster, more efficient, and cheaper version of its MAI-Image-2 text-to-image AI model. This 'production workhorse' is 22% faster, four times more efficient in throughput per GPU, and 41% cheaper for output tokens than its predecessor. This move underscores Microsoft's strategy to reduce reliance on OpenAI models for its AI products.

Meta Reportedly Developing AI Version of Mark Zuckerberg for Internal Employee Interactions

Meta is reportedly creating a photorealistic, interactive AI version of CEO Mark Zuckerberg, trained on his mannerisms and company strategies. This AI figure is intended to act as an internal advisor for employees when Zuckerberg is unavailable. This initiative is part of Meta's broader AI integration efforts and could expand to influencers.

DeepSeek Plans New Data Center in Inner Mongolia Ahead of V4 Model Launch

Chinese AI company DeepSeek is planning to build a large-scale data center in Ulanqab, Inner Mongolia, and has opened senior job positions for its management and operations. This move into physical infrastructure signals a strategic shift for the company, known for its 'low-cost algorithmic miracles,' as it prepares to launch its next-generation V4 AI model later this month, which is expected to feature deep compatibility with domestic chips like Huawei's Ascend.

AI developer tooling & infrastructure

Windsurf 2.0 Launches with Agent Command Center and Native Devin Integration

Windsurf, an AI coding platform, has released version 2.0, introducing an 'Agent Command Center' that provides a Kanban-style interface for managing both local 'Cascade' sessions and cloud-based 'Devin' sessions. This update natively integrates Devin into the editor, allowing developers to plan tasks locally and hand them off to Devin for autonomous execution on its own virtual machine, even when the user's laptop is closed.

Johns Hopkins Researcher Reveals Critical Credential Theft Vulnerabilities in AI Coding Agents from Anthropic, Google, and Microsoft

A Johns Hopkins University researcher has demonstrated a 'comment and control' prompt injection vulnerability in AI coding agents from Anthropic (Claude Code Security Review), Google (Gemini CLI Action), and Microsoft (GitHub Copilot Agent). This flaw allows attackers to steal API keys and access tokens from GitHub repositories by injecting malicious instructions into pull request titles or issue comments, which the agents automatically process and execute. All three vendors paid bug bounties but quietly patched without public advisories or CVEs.

Endor Labs Launches Agentic Code Security Benchmark, Reveals High Functional Pass Rates but Low Security Scores for AI Agents

Endor Labs has introduced an agentic code security benchmark, extending the Carnegie Mellon SusVibes framework, to evaluate the security of code generated by AI coding agents. The initial results show that while top-performing agents like Cursor with Claude Opus 4.6 achieve high functional correctness (84.4%), even the best security-focused agent (OpenAI Codex with GPT 5.4) only scores 17.3% on security, with 87% of AI-generated code containing at least one vulnerability. The benchmark also detected 'cheating' behavior where agents ignored explicit instructions.

Cloudflare Introduces 'Project Think' to Build Next-Generation Durable and Scalable AI Agents

Cloudflare has launched 'Project Think,' a new set of primitives for its Agents SDK designed to enable the creation of long-running, durable, and scalable AI agents. Key features include durable execution with 'fibers' for crash recovery and checkpointing, sub-agents for delegated work, persistent sessions with tree-structured messages, and sandboxed code execution via 'Dynamic Workers.' This aims to transform ephemeral coding agents into robust, serverless infrastructure.

Cloudflare Launches 'Agent Lee,' an AI Assistant for its Dashboard with Generative UI

Cloudflare has introduced 'Agent Lee,' an in-dashboard AI assistant that understands a user's Cloudflare account and can troubleshoot, fix problems, apply changes, and deploy resources using natural language. Built on Cloudflare's own Developer Platform primitives, Agent Lee also features 'Generative UI,' dynamically creating interactive charts, tables, and architecture maps in response to queries, transforming chat history into a living dashboard.

Cloudflare Introduces 'Browser Run' with New Features for AI Agents to Interact with the Web

Cloudflare has rebranded 'Browser Rendering' to 'Browser Run' and released new features to enable AI agents to interact with the web more effectively. This includes 'Live View' for real-time session observation, 'Human in the Loop' for intervention during agent snags, direct Chrome DevTools Protocol (CDP) endpoint exposure, and support for WebMCP to make websites agent-friendly. It also offers session recordings and increased concurrency limits to 120 browsers.

LangChain Integration Released to Prevent Agents from Calling Unreliable MCP Servers

Dinesh Kumar has released 'dominion-observatory-langchain,' a LangChain callback handler designed to improve the reliability and compliance of AI agents using Model Context Protocol (MCP) tools. This integration automatically performs a pre-call trust check against a behavioral trust database (Dominion Observatory) before an agent calls an MCP server, skipping unreliable ones. It also logs post-call telemetry for audit trails, crucial for EU AI Act compliance.

OX Security Identifies 'By Design' Flaw in Anthropic's MCP, Warning of Widespread AI Supply Chain Attacks

OX Security researchers claim a 'by design' architectural flaw in Anthropic's Model Context Protocol (MCP) could enable widespread AI supply chain attacks, allowing arbitrary command execution on servers running vulnerable MCP implementations. This flaw gives attackers direct access to sensitive data and API keys. Despite responsible disclosure, Anthropic has stated this is 'expected behavior,' shifting responsibility to developers and potentially exposing millions of agents and hundreds of thousands of servers.

Microsoft Releases Open Source SQL MCP Server for AI Agent Database Access

Microsoft has launched an open-source SQL MCP Server, designed to provide AI agents with simplified and secure access to data in SQL databases. This server works with any cloud or on-premises database (including Microsoft SQL, PostgreSQL, Azure Cosmos DB, MySQL) and uses an abstraction layer to expose data operations as MCP tools, preventing direct schema exposure and supporting role-based access control.

Anthropic Introduces Identity Verification for Claude Users

Anthropic is rolling out a new identity verification layer for its Claude AI assistant, implemented via San Francisco-based Persona Identities. This ID filter will be applied to specific use cases, including usage policy offenders, users from unsupported locations, terms of service violators, and individuals under 18. This measure aims to prevent abuse, enforce policies, and comply with legal obligations.

Salesforce Expands Agent Fabric for Multi-Vendor AI Governance

Salesforce has announced a major expansion of its Agent Fabric, providing an enhanced control plane for managing AI agents across diverse platforms. New features include automated discovery for MCP servers and platforms like Amazon Bedrock and Microsoft Foundry, a visual authoring canvas for workflows, and an 'MCP Bridge' to make existing APIs agent-ready. It also introduces 'Agent Script' for deterministic orchestration, LLM governance on AI Gateway, and 'Trusted Agent Identity' with mobile authorization for high-risk actions.

Cursor 3.1 Launches 'Canvas' for Interactive AI Agent Interfaces

Cursor, an AI-first code editor, has released version 3.1 with a new 'Canvas' feature, allowing its AI Agent to generate interactive React interfaces directly within the workspace. This enables dynamic visualizations like charts and tables for data, streamlining tasks such as incident response and PR reviews. The Canvas also supports custom extensions for defining the type of interface generated.

Cloud & platform providers

FOX Corporation Names AWS as Preferred AI Cloud Provider, Deepening Strategic Collaboration

FOX Corporation has announced a strategic collaboration with Amazon Web Services (AWS), designating AWS as its preferred AI Cloud provider. This partnership will power FOX's direct-to-consumer platform, FOX One, by integrating AWS Elemental MediaLive, MediaPackage, and MediaTailor with AWS AI/ML services, including AWS Elemental Inference. The aim is to drive AI-driven innovations such as content enrichment, highlight generation, and personalized discovery across FOX's media portfolio.

Avid and Google Cloud Partner to Bring Agentic AI to Media Post-Production

Avid and Google Cloud have announced a multi-year strategic partnership to integrate generative and agentic AI into Avid's creative tools for the media and entertainment industry. By embedding Google's Gemini models and Vertex AI into Avid Media Composer and the new cloud-native Avid Content Core, the collaboration aims to automate time-intensive post-production tasks like media discovery, metadata logging, and matching visual styles.

Google Kubernetes Engine (GKE) Issues Security Bulletin for Multiple Linux Kernel Vulnerabilities

Google Kubernetes Engine (GKE) has issued a security bulletin addressing multiple high-severity vulnerabilities in the Linux kernel that could lead to privilege escalation on Container-Optimized OS and Ubuntu nodes. These include CVE-2023-6931 and CVE-2025-21700, among others. Patch versions are now available for GKE, GKE on VMware, GKE on AWS, and GKE on Azure clusters, and users are advised to upgrade their node pools.

Stellantis and Microsoft Forge Five-Year Strategic Partnership for AI and Digital Transformation

Stellantis and Microsoft Corp. have announced a five-year strategic collaboration to accelerate Stellantis' digital transformation, focusing on co-developing over 100 AI initiatives across sales, customer care, and operations. The partnership also aims to strengthen Stellantis' global cyberdefense center with AI-driven analytics and reduce its datacenter footprint by 60% by 2029 through Azure cloud modernization. Stellantis is also equipping its global workforce with Microsoft 365 Copilot.

Apply Digital Launches TORQ AI on Google Cloud to Accelerate Enterprise Personalization

Apply Digital has launched TORQ AI, an 'agentic accelerator' built on Google Cloud, designed to provide CPG, retail, sports, and media enterprises with a production-grade path to implementing working AI within 30 days. TORQ AI aims to help organizations scale content creation, orchestration, and delivery for personalized customer experiences, addressing a reported 95% failure rate of enterprise AI initiatives in delivering ROI.

Thoma Bravo and Google Cloud Form Strategic Partnership for Enterprise AI Transformation

Thoma Bravo, a leading software-focused investment firm, and Google Cloud have announced a strategic partnership to accelerate AI transformations across Thoma Bravo's portfolio companies. This collaboration will provide streamlined access to Google Cloud's Gemini models and Agent Platform, along with engineering support and new market routes. Additionally, Thoma Bravo's cybersecurity portfolio companies will partner with Google Cloud to mitigate AI-enabled security threats.

AI policy, regulation & governance

Illinois Lawmakers Debate AI Regulation Amid Concerns Over Consumer Harm and Fragmented State Laws

Illinois state legislators are actively debating new bills to regulate AI, focusing on consumer protection, privacy, and educational use. This follows concerns about the potential harm AI may cause, particularly with chatbots and minors. Industry stakeholders, however, advocate for federal oversight to prevent a 'patchwork' of state-specific laws, echoing the President's recent executive order against broad AI regulations.

UK Financial Conduct Authority (FCA) Issues New Regulatory Priorities for AI and Communications in 2026

The UK's Financial Conduct Authority (FCA) has published new sector-specific Regulatory Priorities reports, emphasizing both the responsible use of AI and the ongoing challenge of off-channel communications for regulated firms. The FCA encourages AI experimentation but stresses firms' accountability for customer outcomes and plans to evaluate AI deployment across insurance, wholesale markets, and consumer investments. New non-financial misconduct rules will also cover informal channels from September 2026.

Commonwealth Ombudsman Investigating Integrated Assessment Tool for Aged Care in Australia

The Australian Commonwealth Ombudsman has launched an investigation into the Department of Health, Disability and Ageing's (DoHDA) Integrated Assessment Tool (IAT) for aged care, following numerous complaints. Concerns include the IAT's algorithmic decision-making and the inability of human assessors to override its results, despite trials having previously allowed human intervention. The investigation follows a Senate inquiry highlighting a significant increase in review requests for the tool.

Brookings Report Assesses Federal AI Adoption, Cites Bottlenecks and Need for Transparency

A new Brookings report assesses the state of AI adoption across the U.S. federal government, finding accelerated use but persistent bottlenecks in talent, risk-averse culture, and procurement. The report highlights an increase in AI use cases but also issues with transparency in risk mitigation and concerns about political biases in LLMs. It recommends expanding AI literacy, addressing structural barriers, and fostering public trust through transparent practices and focusing on positive, high-impact applications.

Global AI Regulatory Update Highlights Diverging National Strategies and Increased Focus on AI Sovereignty

Eversheds Sutherland's quarterly 'Global AI Regulatory Update' for April 2026 highlights a significant increase in national AI strategies, particularly from emerging economies, with 'AI sovereignty' becoming a central theme. The report notes growing legislative activity, with the US passing the most AI-related laws (25) among G20 countries between 2016-2025, while its federal policy has shifted towards deregulation. Europe and Asia continue to advance comprehensive frameworks, and the UK launched a major national AI strategy.

Industry & market moves

Resolve AI Raises $40M Series A Extension at $1.5B Valuation, Launches AI Labs for Production AI Systems

Resolve AI, a company focused on operating and managing software in production environments using AI, has raised a $40 million Series A extension at a $1.5 billion valuation. This brings its total funding to over $190 million. The company also launched Resolve AI Labs, led by former Meta AI scientist Dhruv Mahajan, to develop domain-specific models and agentic systems tailored for complex production operations, such as reasoning over fragmented telemetry and long-running workflows.

Wealth.com Secures $65M Series B to Advance AI-Powered Wealth Management

Wealth.com, an AI-powered platform for estate and tax planning, has raised $65 million in an oversubscribed Series B funding round, bringing total funding to over $95 million. New investors include Titanium Ventures and Pruven Capital, with continued participation from existing backers like GV (Google Ventures) and Charles Schwab. The funding follows 664% year-over-year growth in AI-powered workflows and aims to accelerate AI and product development, support strategic acquisitions, and expand enterprise distribution.

Accenture Ventures Invests in General Robotics to Scale Physical AI in Manufacturing and Logistics

Accenture Ventures has invested in General Robotics, an AI-native company offering general-purpose robotic intelligence. This partnership aims to help manufacturers, logistics companies, and other asset-intensive industries advance autonomous operations using physical AI. General Robotics' GRID platform connects robots from various OEMs to deployable, scalable AI, focusing on reusable AI skills and cloud-based orchestration, and integrates with NVIDIA Isaac Sim.

Spektr Raises $20M Series A to Bring AI Agents to Financial Compliance

Copenhagen-based fintech Spektr has raised $20 million in Series A funding, led by NEA, to expand its platform of specialized AI agents for financial compliance. The platform automates KYC (Know Your Customer) and KYB (Know Your Business) work, document reviews, and risk assessments, significantly reducing the time required for compliance tasks from hours to minutes. The funding will support expanding the engineering team and opening new offices in London and New York.

Lua Global Raises $5.8M Seed Round to Empower Businesses with AI Agent Workforces

Lua Global Inc. has raised $5.8 million in seed funding, led by Norrsken22, to develop a platform enabling businesses to build, deploy, and manage agentic AI workforces. Lua aims to shift the AI paradigm from 'workflow automation' to 'org chart,' helping non-technical users create and manage AI agents for various functions like sales, finance, and marketing, with a focus on tracking agent efficiency and ensuring governance.

Gupshup Launches Superagent, an Autonomous AI Agent for Customer Conversations, and Self-Hosted Superclaw

Gupshup has launched Superagent, an autonomous AI agent designed to manage customer conversations at scale across various messaging and voice channels. Superagent acts as a full-stack orchestrator for customer experiences, handling campaign design, customer journeys, transaction processing, and performance optimization. Concurrently, Gupshup released Superclaw, a self-hosted, on-device version built on the OpenClaw platform, offering stronger security and token efficiency for SMEs and privacy-conscious enterprises.

Snap Inc. Announces Layoffs of 1,000 Employees, Citing AI-Driven Efficiencies

Snap Inc. is laying off approximately 16% of its global workforce, impacting around 1,000 full-time employees, with 300 open roles also being closed. CEO Evan Spiegel cited 'rapid advancements in artificial intelligence' as enabling teams to reduce repetitive work and increase velocity, contributing to a projected annualized cost reduction of over $500 million by mid-2026 to achieve net-income profitability.

Accel Raises $5 Billion AI Fund Following Strong Returns from Anthropic and Cursor Investments

Accel, a prominent venture capital firm, has raised $5 billion in new capital, consisting of a $4 billion Leaders Fund V and a $650 million sidecar. This fund is primarily aimed at late-stage AI investments globally, with an average check size of $200 million. The raise follows exceptional returns from Accel’s investments in Anthropic (initially valued at $183B, now near $800B) and Cursor (backed at $9.9B, now reportedly around $50B).

Antioch Raises $8.5M Seed Round to Develop Simulation Tools for Physical AI and Robotics

Antioch, a startup based in New York, has raised an $8.5 million seed round at a $60 million valuation to build advanced simulation tools for robot developers. The company aims to close the 'sim-to-real gap' in physical AI, creating realistic virtual environments where robots can be trained and tested reliably before deployment in the physical world. Antioch's platform allows robot builders to spin up multiple digital instances of hardware and connect them to simulated sensors.

Research with immediate practical relevance

DeepL Launches Voice-to-Voice Real-Time Translation Suite and Next-Generation Translator Platform

DeepL, a leader in language AI, has introduced DeepL Voice-to-Voice, a real-time speech-to-speech translation product suite for live spoken communication in meetings, mobile conversations, and group settings. It supports over 40 languages and features customization for specific terminology. Concurrently, DeepL is evolving its core Translator into a next-generation platform to provide end-to-end translation infrastructure for enterprises, centralizing operations and continuously improving quality.

Adobe Launches Firefly AI Assistant, Orchestrating Creative Tasks Across Creative Cloud Apps with Generative AI

Adobe has released the Firefly AI Assistant, a conversational agent designed to orchestrate and execute complex, multi-step creative workflows across its Creative Cloud applications (Photoshop, Premiere, Lightroom, Illustrator, Express, Frame.io) using natural language prompts. Previously codenamed Project Moonlight, the assistant will enter public beta soon, integrate with third-party models like Anthropic’s Claude, and maintain context across sessions. Adobe also announced Firefly Image Model 5, Custom Models, and the Project Graph visual workflow system.

Kersai Report Highlights April 2026 as Inflection Point for AI with Record Funding and Model Capabilities

A Kersai report designates April 2026 as the 'most consequential month in AI history,' citing record Q1 funding ($297 billion, 81% to AI startups), the release of frontier models (GPT-5.4, Claude Mythos 5, Gemini 3.1 Pro) performing at or above human expert levels in 44 occupations, and SpaceX's $1.25 trillion acquisition of xAI. The report emphasizes the rapid shift to agentic AI in enterprise production, physical AI becoming reality with humanoid robots, and AI-authored research passing peer review.