The Informed Layer · Newsscan
Daily briefing

Anthropic soars; AI agents drive infra, security, law

← All issues

Anthropic / Claude ecosystem

Anthropic Debuts Claude Code Desktop App Redesign and 'Routines' for Automated Workflows

Anthropic has released a significantly redesigned Claude Code desktop app, featuring a new sidebar for parallel sessions, an integrated terminal, and an in-app file editor. Concurrently, they introduced 'Routines' in research preview, enabling users to configure Claude Code automations to run on a schedule, via API calls, or in response to GitHub events on Anthropic's web infrastructure, even when the user's laptop is closed. These routines are available to Pro, Max, Team, and Enterprise plan users with daily usage limits.

Anthropic's Mythos AI Model Sparks Urgent Discussions in Washington and Wall Street

Anthropic's new Mythos AI model, touted for its unprecedented cybersecurity capabilities in finding software vulnerabilities, has led to urgent discussions among US Treasury Secretary Scott Bessent, Federal Reserve Chair Jerome Powell, and top Wall Street executives. The model, released to a very limited number of carefully selected parties as part of 'Project Glasswing,' is being tested by federal agencies, despite a previous ban on working with Anthropic, and is raising alarms among UK financial regulators.

Federal Agencies Reportedly Skirt Trump's Anthropic Ban to Test Mythos AI

Politico reports that federal agencies and government officials are quietly sidestepping President Donald Trump’s ban on working with Anthropic to test its advanced AI model, Mythos. The Commerce Department's Center for AI Standards and Innovation is actively testing Mythos’ hacking prowess, and Anthropic co-founder Jack Clark confirmed the company has briefed the Trump administration on the model's capabilities, despite a previous contract dispute with the Pentagon.

Frontier model providers

OpenAI Unveils GPT-5.4-Cyber for Defensive Cybersecurity

OpenAI has announced GPT-5.4-Cyber, a specialized variant of its latest flagship model, GPT-5.4, fine-tuned specifically for defensive cybersecurity workflows. Access to this 'cyber-permissive' model is limited to vetted security vendors, organizations, and researchers through an expanded 'Trusted Access for Cyber' program. This model offers advanced capabilities such as binary reverse engineering for analyzing compiled software for malware and vulnerabilities.

Google DeepMind Releases Gemini Robotics-ER 1.6 with Enhanced Embodied Reasoning

Google DeepMind has released Gemini Robotics-ER 1.6, a significant upgrade to its embodied reasoning model, designed to be the 'cognitive brain' for robots in real-world environments. This model introduces enhanced spatial and physical reasoning, improved multi-view understanding, and a new 'instrument reading' capability to interpret analog and digital gauges in industrial settings. Boston Dynamics has already integrated this model into its Spot inspection platform.

Microsoft Unveils MAI Image 2 Efficient AI Model: A 'Production Workhorse'

Microsoft has introduced MAI Image 2 Efficient, a new version of its text-to-image AI model, which is 22% faster, four times more efficient, and 41% cheaper than its predecessor. Described as a 'production workhorse,' this model is designed for large-scale tasks such as generating product images, marketing visuals, and UI mockups in batches, offering 'production-ready quality.'

Meta Expands Custom AI Chip Partnership with Broadcom to Power Multi-Gigawatt AI Ecosystem

Meta has significantly expanded its partnership with Broadcom to co-develop multiple generations of custom AI silicon ('XPUs') for its MTIA (Meta Training and Inference Accelerator) lineup, extending the deal until 2029. The collaboration includes an initial commitment exceeding one gigawatt of computing capacity, with plans for a multi-gigawatt rollout to power AI across Meta's applications and services, reducing reliance on general-purpose GPUs.

DeepSeek V4 to Launch Late April, Reportedly Transitioning to Huawei Ascend Chips

DeepSeek's next-generation flagship model, DeepSeek V4, is expected to launch in late April 2026. Reports suggest it will feature a total parameter count nearing 1 trillion with a 1 million token context window, and crucially, will run entirely on Huawei Ascend 950 PR chips, shifting away from NVIDIA GPUs and the CUDA ecosystem. This move is seen as a key milestone for China's AI industry in reducing reliance on US hardware.

AI developer tooling & infrastructure

Cloudflare Expands Agent Cloud with Dynamic Workers, Sandboxes, and Git-Compatible Storage for Production AI Agents

Cloudflare has significantly expanded its Agent Cloud with new features to help developers build, deploy, and scale AI agents from experimental demos to robust, production-grade workloads. Key additions include 'Dynamic Workers,' an isolate-based runtime for efficient AI-generated code execution; 'Artifacts,' Git-compatible storage for agent-scale repositories; and the general availability of 'Sandboxes,' providing persistent, isolated Linux environments for agents. The platform also integrates OpenAI's GPT-5.4 and Codex models.

Apache Airflow Introduces Common AI Provider with Native LLM and AI Agent Support

Apache Airflow has launched the 'Common AI Provider' (apache-airflow-providers-common-ai 0.1.0), which natively integrates LLM and AI agent capabilities directly into Airflow workflows. This new provider package, built on Pydantic AI, offers 6 operators, 6 TaskFlow decorators, and 5 toolsets, supporting over 20 model providers including OpenAI, Anthropic, Google, and AWS Bedrock. It features durable=True for caching agent steps on retry and built-in human-in-the-loop (HITL) review mechanisms.

Microsoft Copilot to Integrate Autonomous Background Agents with UI-Based Automation

Microsoft is fundamentally shifting its Copilot strategy towards autonomous AI agents capable of executing complex tasks 24/7 in the background within the Microsoft 365 ecosystem. This proactive AI system, driven by Microsoft Copilot Studio, will monitor data streams and trigger actions without constant human intervention. It's also testing 'UI-based automation' to enable Copilot to interact with legacy software lacking modern APIs, effectively bridging AI with aging enterprise infrastructure.

Palo Alto Networks Acquires Koi to Launch Agentic Endpoint Security (AES)

Palo Alto Networks has completed its acquisition of Koi, defining a new category called Agentic Endpoint Security (AES) to protect 'vibe coding agents' and autonomous endpoint tools. This acquisition integrates Koi's technology with Prisma AIRS and will introduce a new module for Cortex XDR to identify and remediate risks within the AI software ecosystem. It addresses the expanded attack surface created by the rapid adoption of agentic tools like Claude Code and OpenClaw.

Kong AI Gateway Now Supports Agent-to-Agent (A2A) Traffic for Comprehensive AI Governance

Kong Inc. has announced Kong Agent Gateway within its AI Gateway 3.14 release, making it the only gateway solution to support advanced LLM, Model Context Protocol (MCP), and agent-to-agent (A2A) communication. This new capability provides a unified control plane for governing all AI traffic types, enabling unified observability, production-ready agentic AI, cost visibility, and reduced compliance risk across multi-agent workflows.

AppControl Introduces AI-Powered Analysis for Windows PC Performance and Privacy

AppControl has announced a new integration with AI agents and IDEs like Claude, Cursor, Windsurf, and Gemini CLI, enabling natural-language analysis of a Windows PC's recent 72 hours of activity. This update turns its DVR-style system resource monitoring into a conversational tool to diagnose performance and privacy issues, allowing users to query their system's behavior in plain English.

TinyFish Launches 4-in-1 Web Agents API Suite for Developers

TinyFish has expanded its web infrastructure platform into a four-product API suite for developers building production AI coding agent workflows. The suite offers Web Agent for autonomous multi-step execution, Web Search for structured JSON results, Web Browser for managed stealth Chrome sessions, and Web Fetch for converting URLs to clean Markdown/HTML/JSON. These are accessible via a single API key and a CLI/Agent Skill integration.

Cursor 3.1 Released with Parallel Agents and Improved Voice Input

Cursor, an AI-first code editor, has released version 3.1, focusing on an enhanced Agents Window. Key new features include a 'tiled layout' that allows users to run multiple AI agents in parallel within the same window and compare their outputs side-by-side. The update also includes a rewritten voice input system for improved transcription accuracy and the ability to select a specific Git branch when starting a cloud agent.

GitGuardian Report Reveals 29 Million Leaked Secrets in 2025, Blames AI-Assisted Code

GitGuardian's 'State of Secrets Sprawl Report' found 28,649,024 new secrets exposed in public GitHub commits in 2025, a 34% year-over-year increase. The report attributes a significant portion of this surge to AI-assisted code, noting that commits co-authored by Claude Code leaked secrets at roughly double the baseline rate. It highlights how authentication governance is failing to keep pace with the velocity of AI-driven development and multi-provider integrations.

Cloud & platform providers

AWS Interconnect – Multicloud and Last Mile Now Generally Available

AWS has announced the general availability of AWS Interconnect – Multicloud, a managed private connectivity service connecting Amazon VPCs directly to VPCs on other cloud providers (starting with Google Cloud, Azure later in 2026). Additionally, AWS Interconnect – Last Mile simplifies high-speed, private connections from on-premises networks to AWS through existing network providers. Both services are fully managed, offering predictable latency, MACsec encryption, and multi-region resilience.

AWS Launches Amazon Bio Discovery, an AI-Powered Application for Drug Discovery

Amazon has launched Amazon Bio Discovery, an AI-powered application designed to accelerate early-stage drug discovery by enabling scientists to run complex computational workflows without writing code. The platform provides access to a library of specialized biological foundation models and an AI agent to guide model selection, parameter setting, and results interpretation. It integrates with lab partners for wet-lab validation, creating a feedback loop to refine predictions.

AWS Launches Agent Registry to Centralize AI Agent Governance

AWS has launched Agent Registry, a centralized platform for housing, building, and governing AI agents across various frameworks. Available in preview through the AWS Bedrock AgentCore console, it provides a structured and searchable format for tracking agent ownership, protocols, capabilities, and invocation methods. This aims to tackle governance and transparency challenges as enterprises scale AI agents, addressing lack of visibility, weak governance, and tool duplication.

Cloudflare Introduces Mesh for Secure Private Networking Across Users, Nodes, and AI Agents

Cloudflare has launched Cloudflare Mesh, a new private networking solution designed for the rise of AI agents. Mesh unifies AI agents, humans, and multi-cloud infrastructure into a single secure fabric, allowing agents to securely access private databases, internal APIs, and staging environments. It integrates with Cloudflare Developer Platform (Workers, Durable Objects, Agents SDK) and leverages Cloudflare One's security controls, providing a full bidirectional, many-to-many network routed through Cloudflare’s global network.

GitLab Expands Google Cloud Partnership for Vertex AI Integration with Duo Agent Platform

GitLab has expanded its partnership with Google Cloud, allowing customers to use Vertex AI models, including Gemini, with the GitLab Duo Agent Platform. This integration enables AI agents within GitLab to call foundation models through Vertex AI from within GitLab, with usage counting towards existing Google Cloud spending commitments. It connects GitLab's software development and security workflow tools with Google Cloud's managed AI model service.

AI policy, regulation & governance

US Lawyers Warn Clients About AI Chat Confidentiality Following Federal Ruling

Following a federal judge's ruling in New York that a former CEO's AI chats could not be shielded from prosecutors in a securities fraud case, some U.S. lawyers are urgently advising clients not to treat AI chatbots like confidential sources. Conversations with tools like Anthropic's Claude and OpenAI's ChatGPT could be subject to discovery in legal proceedings, as no attorney-client relationship exists with an AI platform.

EU AI Act Sets August 2026 Deadline for AI Recruitment Regulation, US Bias Case Advances

The EU AI Act's full enforcement kicks in on August 2, 2026, classifying all AI systems in recruitment, task allocation, and performance monitoring as 'high-risk,' requiring mandatory risk assessments, bias testing, and human oversight with significant fines for non-compliance. Concurrently, a US federal court has allowed a landmark AI hiring bias class action (Mobley v. Workday) under the Age Discrimination in Employment Act to proceed as a nationwide collective action.

KnowBe4 Launches Agent Risk Manager for Autonomous AI Agent Security

KnowBe4 has introduced Agent Risk Manager, a new security product aimed at organizations using autonomous AI agents in operational workflows. Part of its HRM+ platform, this tool monitors agent behavior in real time, applies behavioral guardrails, detects prompt injection attacks, and identifies unauthorized sharing of sensitive information and unbounded consumption of resources. It aims to address the security gap created by the shift from AI-assisted tasks to more autonomous systems.

Australia Leads in Responsible AI Adoption, Lags in Productivity Gains

A new KPMG report indicates that Australian organizations are taking a 'responsible' approach to AI adoption, leading global counterparts in AI governance (31.6% implementing trustworthy AI policies vs. 26.3% global average) and cybersecurity (37.8% vs. 25.9% global average). However, Australia is falling behind in realizing productivity gains from automation (34.7% vs. 42.3% global average) and leveraging advanced analytics and real-time insights.

Queensland Government to Criminalize AI-Generated 'Deepfake' Sexual Images

The Queensland Government in Australia plans to criminalize the creation or alteration of intimate images of a person without consent, including those generated by artificial intelligence or image-editing tools. Under the proposed laws, offenders could face up to three years in jail. This reform specifically targets the growing issue of AI deepfakes used for cyberbullying and other exploitative purposes, extending beyond traditional image-based abuse laws to cover entirely fabricated content.

Industry & market moves

Anthropic Reportedly Seeks New Funding at $800 Billion Valuation, Eyeing 2026 IPO

Anthropic is reportedly seeking new investment offers at a valuation of around $800 billion, more than double its $380 billion post-money valuation from two months prior. This comes shortly after the introduction of its advanced, but publicly restricted, Mythos AI model. The company is also in early talks with Goldman Sachs, JPMorgan, and Morgan Stanley for a potential IPO as early as October 2026, which bankers believe could raise over $60 billion.

Samsung SDS Shares Jump 20% on KKR Partnership and $820 Million Convertible Bond Purchase

Shares of Samsung SDS surged by nearly 20% after global private equity firm KKR agreed to invest 1.22 trillion won ($820 million) in newly issued convertible bonds. The deal will back Samsung SDS's expansion into AI infrastructure and global markets as a full-stack AI solutions provider. KKR will also take an advisory role on strategic areas including M&A, capital allocation, and AI offerings.

OpenAI Acquires Personal Finance AI Startup Hiro Finance

OpenAI has acquired Hiro Finance, an AI-powered personal finance planning startup, with founder Ethan Bloch and his staff joining OpenAI. The terms of the acquisition were not disclosed. Hiro's product, which offered AI-powered financial planning and scenario modeling, will shut down on April 20, with all user data deleted by May 13. This marks OpenAI's second acquisition in the financial services space.

Novo Nordisk Partners with OpenAI for Enterprise-Wide AI Integration in Drug Discovery

Novo Nordisk has announced a strategic partnership with OpenAI to integrate AI capabilities across its global operations, from drug discovery and manufacturing to supply chain and commercial functions. The collaboration aims to accelerate scientific discovery, optimize processes, and bring new therapies to market faster by leveraging OpenAI's advanced AI tools and expertise, with full integration targeted by the end of 2026. Financial terms were not disclosed.

Camtek Acquires Visual Layer to Enhance Visual AI Capabilities in Semiconductor Inspection

Camtek Ltd., a developer of inspection and metrology equipment for the semiconductor industry, has signed a definitive agreement to acquire Visual Layer, a Tel Aviv-based AI company specializing in visual analytics. This acquisition aims to deepen Camtek's visual AI capabilities, integrating Visual Layer's proprietary technology and skilled team to address the increasingly demanding requirements of the semiconductor market, particularly in improving product throughput and performance.

Research with immediate practical relevance

NVIDIA Launches Ising: First Open-Source Quantum AI Models for Scalable Quantum Computing

NVIDIA has announced 'Ising,' the world’s first family of open-source quantum AI models, designed to accelerate the development of useful quantum computers. Ising includes two main models: 'Ising Calibration,' a vision-language model for automating quantum processor calibration from days to hours, and 'Ising Decoding,' two 3D CNN variants for real-time quantum error correction that are up to 2.5x faster and 3x more accurate than existing standards. These models are available on Hugging Face and GitHub.