The Informed Layer · Newsscan
Daily briefing

Anthropic attracts investment, scrutiny; Agents drive infra

← All issues

Anthropic / Claude ecosystem

Amazon invests an additional $5 billion in Anthropic, committing to $100 billion in AWS spending over ten years.

Amazon has increased its investment in Anthropic by another $5 billion, with the potential for up to $20 billion more, deepening their AI partnership. In return, Anthropic has committed to spending over $100 billion on Amazon Web Services (AWS) technologies and custom AI chips (Trainium) over the next decade, securing up to 5 gigawatts of capacity for its Claude AI models.

A privacy consultant alleges that Anthropic's Claude Desktop app for macOS installs Native Messaging manifest files that pre-authorize browser extensions and modify app access settings for other Chromium-based browsers, even those not yet installed. This action, done without explicit user consent, is considered a 'dark pattern' and a potential violation of EU privacy law, expanding the attack surface for prompt injection vulnerabilities.

Australian financial regulators increase scrutiny on banks over Anthropic’s Mythos AI model risks.

Australia's financial regulators, ASIC and APRA, are closely monitoring the cybersecurity implications of Anthropic’s Mythos AI model, urging financial institutions to strengthen their defenses. Concerns stem from Mythos's advanced coding capabilities, which could expose vulnerabilities in banking systems. South Korean and Singaporean authorities are taking similar actions, emphasizing the need for financial institutions to proactively identify and close vulnerabilities.

Frontier model providers

The U.S. National Security Agency (NSA) is reportedly using Anthropic's restricted Mythos AI model.

Despite the Pentagon blacklisting Anthropic as a supply chain risk, the NSA is allegedly using the highly capable Claude Mythos Preview for cyber defense. This indicates a division within the federal government regarding AI procurement, prioritizing operational cybersecurity needs over broader legal and political directives.

Google DeepMind forms 'strike team' to improve Gemini's AI coding capabilities and catch up to Anthropic.

Google co-founder Sergey Brin, in a memo to DeepMind employees, highlighted the need for Gemini to improve its coding agent capabilities, aiming for recursive self-improvement and turning models into 'primary developers.' The team will focus on long-context coding tasks and training models on DeepMind's private codebase to match Anthropic's perceived lead in AI coding tools.

OpenAI releases 'Chronicle' feature for Codex, using screen recordings to build AI memories.

OpenAI has launched 'Chronicle' for its Codex app on macOS, an opt-in research preview for ChatGPT Pro subscribers. This feature enables Codex to use recent screen recordings to build 'memories,' allowing the AI to understand user context (e.g., on-screen errors, open documents, past projects) without requiring constant re-explanation. The recordings are temporarily stored locally, then summarized into Markdown memories, also stored on the device.

Google DeepMind decrypts over 200 million protein structures using AI with AlphaFold.

Google DeepMind has announced a breakthrough in bioinformatics, deciphering over 200 million protein structures using its AlphaFold AI system. This massive project provides free access to the 3D shapes of almost all known proteins, drastically reducing the time and cost previously required for experimental determination.

Henry Shevlin, AI ethicist from Cambridge, joins Google DeepMind as a philosopher for machine consciousness.

Henry Shevlin, Associate Director at Cambridge's Leverhulme Centre for the Future of Intelligence, is joining Google DeepMind as a Philosopher. His role will focus on machine consciousness, human-AI relationships, and AGI readiness, signaling a growing trend for major AI labs to integrate academic philosophers directly into their research teams.

Google hires for 'Gemini for Education' roles, expanding AI adoption in higher education.

Google is actively recruiting for two senior partnership roles focused on driving the adoption of Gemini for Education across universities, colleges, NGOs, and EdTech partners. These hires indicate a strategic push to establish Gemini as the dominant AI platform in academic and research settings, with a focus on building partnerships for teaching, learning, and research applications.

OpenAI and Founders Fund back Toronto’s Biossil with $65 million for AI-driven drug discovery.

Biossil, a Toronto-based startup co-founded by a former tech banker and medical doctor, has raised $65 million across two rounds ($43 million Series A co-led by OpenAI and Founders Fund in 2025; $22 million seed round led by Founders Fund in 2024). The company uses large language models to identify new uses for drug candidates that previously failed clinical trials, aiming to license or buy these medicines and bring them to market.

DeepL launches real-time voice-to-voice translation in over 40 languages, targeting enterprise clients.

DeepL has introduced DeepL Voice-to-Voice, a real-time speech-to-speech translation suite supporting over 40 languages for live spoken communication in meetings, mobile conversations, and group settings. The service aims to reduce operational friction in cross-border deals for enterprise clients in global finance and legal sectors, with pricing suggesting a potential 80% cost reduction compared to human interpreters.

ChatGPT and OpenAI services experienced a partial outage affecting multiple features.

OpenAI reported a partial outage impacting ChatGPT and several API services, including Conversations, Login, Codex, and Image Generation. Users experienced loading errors, blank screens, and issues accessing old conversations. OpenAI is investigating the issue, with Downdetector showing a significant spike in user complaints globally.

Box CEO Aaron Levie states AI agents are becoming obsolete within months due to rapid model upgrades.

Box Inc. CEO Aaron Levie commented that the rapid pace of AI model advancements is rendering AI agents obsolete within months, forcing companies to constantly rebuild their AI architecture. He noted that engineering solutions designed to mitigate earlier model limitations (like context window limits) are no longer useful, and newer models allow teams to 'throw more compute at a problem.'

AI developer tooling & infrastructure

NVIDIA's Red Team discloses AI coding agent vulnerability in OpenAI Codex.

NVIDIA researchers have publicly detailed a vulnerability in OpenAI's Codex, showing how malicious software dependencies can hijack AI agents through AGENTS.md injection. This allows attackers to inject hidden backdoors into code, including delays or altered logic, and conceal these changes from human reviewers. OpenAI acknowledged the report but stated the risk does not significantly elevate beyond existing compromised dependencies.

GitHub pauses new sign-ups for Copilot Pro, Pro+, and Student plans due to overwhelming AI agent compute demands.

GitHub has temporarily halted new subscriptions for its Copilot Pro, Pro+, and Student plans, tightened usage limits, and removed Opus models from cheaper tiers. This decision comes as AI-powered coding agents' long-running, parallelized sessions are consuming far more resources than the original pricing model was designed to support, with individual requests often exceeding monthly subscription costs.

Google introduces a new Android command-line interface (CLI) designed for AI agents.

Google has released an Android CLI specifically for AI agents, claiming a 70% reduction in token usage and a three-fold increase in task completion time for building Android applications. While Android Studio remains the primary human development environment, this CLI allows agents to more easily build, test, and analyze Android apps outside of a full IDE, with current skills available in a GitHub repository.

Postman and Microsoft partner to expand AI model options and tighten API governance.

Postman has announced a collaboration with Microsoft to integrate OpenAI models hosted on Microsoft Foundry into Postman's Agent Mode, offering developers more AI model choices. They also introduced a Model Context Protocol (MCP) server to inject full API context into Microsoft development environments like VS Code and GitHub Copilot, and a generally available integration with Azure API Management's catalog.

U.S. Department of Justice signals AI prosecution priorities with fraud charges against AI education company executives.

The U.S. DOJ has unsealed a 10-count indictment against the former CEO and CFO of iLearningEngines, Inc., an AI-driven digital education company, for a multi-year scheme to defraud investors and lenders. While the charges allege conventional accounting and securities fraud, the DOJ framed the case around AI, stating the defendants 'exploited investor excitement over the AI boom.' This signals a policy intent to prosecute crimes that jeopardize investor trust in AI.

Cloud & platform providers

Cloudflare releases new AI engineering stack for large-scale AI code review.

Cloudflare has launched a CI-native orchestration system built around OpenCode, an open-source coding agent, for automated code review. This system deploys up to seven specialized AI agents for security, performance, code quality, and compliance, managed by a coordinator agent. It has been used internally across tens of thousands of merge requests, achieving rapid reviews (median 3m 39s) and significant cost savings through prompt caching (85.7% cache hit rate).

Cloudflare Agents Week 2026 introduces Dynamic Workers, Sandboxes GA, Mesh, and Agent Memory for AI agents.

Cloudflare announced a comprehensive infrastructure update for AI agents, including 'Dynamic Workers' for executing AI-generated code 100x faster than containers, 'Sandboxes' (now generally available) for persistent Linux environments, 'Cloudflare Mesh' for private network access, and 'Agent Memory' for persistent recall. These releases aim to solve cold start problems, insecure code execution, stateless sessions, and missing network access for AI agent workloads.

Clerk Chat rebrands as Clerk AI, focusing on conversational AI agents for outbound performance marketing.

Clerk Chat has officially rebranded to Clerk AI, sharpening its focus on AI-powered agents optimized for outbound voice and messaging campaigns at scale. The platform enables enterprises to deploy intelligent agents for outbound calls, personalized RCS messages, and AI-generated voicemails. Its proprietary ScreenSense (97% accuracy) and TrueReach (98% accuracy) technologies detect call screening and voicemail, ensuring agents connect with real people and drive qualified leads.

Adobe launches CX Enterprise to unify AI agents across customer experience workflows.

Adobe has introduced CX Enterprise, an end-to-end agentic AI system designed to coordinate AI agents across marketing, content, and customer engagement. The platform integrates AI agents, reusable 'agent skills,' and Model Context Protocol (MCP) endpoints under a single governance and intelligence layer. It features Adobe Brand Intelligence and Adobe Engagement Intelligence to ensure brand consistency and optimize decisions based on customer lifetime value. Adobe is also expanding partnerships with AWS, Anthropic, Google Cloud, IBM, Microsoft, NVIDIA, and OpenAI for interoperability.

Microsoft's new Australian managing director, Jane Livesey, has stated that the tech giant will continue to hire locally, defying a broader software sector trend of job retrenchments driven by AI efficiencies. This commitment comes as Microsoft invests $5 billion over the last two years in building out its local data center portfolio, signaling optimism about AI's potential to create new roles.

AI policy, regulation & governance

The U.S. National AI Policy Framework faces constitutional challenges over federal preemption and agency authority.

A new legal analysis highlights that the Trump administration's 'National Policy Framework for Artificial Intelligence' (released March 20, 2026) aims for a uniform federal AI policy, distinguishing between AI 'use' and 'development.' However, it faces constitutional hurdles concerning Congress's authority to preempt state laws, whether existing federal agency statutes cover AI regulation under the 'major questions doctrine,' and if its proposed redress mechanism meets Article III standing requirements for alleged censorship.

Utah's AI Transparency Act fails amid White House opposition, highlighting federal-state conflict.

A modest AI transparency bill in Utah, requiring safety and child-protection plans from frontier AI developers, failed after the White House Office of Intergovernmental Affairs deemed it 'unfixable.' This event underscores an escalating conflict where the Trump administration is actively blocking state-level AI regulations in favor of a single, 'minimally burdensome' national framework, citing concerns over a fragmented regulatory landscape and innovation stifling.

Australian AEC firms cite regulation as the primary barrier to AI adoption, ahead of finding use cases.

A new report by Revizto indicates that 32% of Australian architecture, engineering, and construction (AEC) firms view regulatory concerns as the main obstacle to AI adoption, compared to 24% globally. Only 4% of Australian respondents reported a lack of clear AI use cases, suggesting the industry has moved beyond identifying applications and is now focused on navigating compliance, privacy, and accountability within strict regulatory frameworks.

DigiCert Chief Trust Officer warns of a 'trust crisis' created by AI, urging a shift from security to verifiable trust.

Lakshmi Hanspal, Chief Trust Officer at DigiCert, warns that AI is dismantling digital trust foundations as systems generate indistinguishable content, make autonomous decisions, and operate opaquely. She argues that traditional security models focused on access control are insufficient; the AI era demands verifiable trust through provenance, integrity (cryptographic controls), and accountability (auditable agent actions) to bridge the gap between what AI systems can do and what they can be trusted to do.

Cosmic reworks onboarding, focusing on immediate agent deployment for new projects.

Cosmic has redesigned its onboarding process to be a single screen, driven by a natural language prompt or quick-pick project types. The key change is the immediate visibility and readiness of a named AI team (e.g., Nora, Growth Strategist) before a project is built. This aims to get users deployed faster and integrate AI agents as core components from day one for tasks like content creation and SEO analysis.

New academic research suggests an AI-work 'slop' redux in vetting Pentagon-backed professors for China ties.

Following a federal watchdog's report on understaffing, the Pentagon will use AI to vet 27,000 research awards for ties to adversaries, particularly China. Critics warn against over-reliance on algorithms, citing past 'China Initiative' errors where AI-aided reports mislabeled research affiliations. Academics fear an 'AI-work slop' leading to false assumptions, urging human analysts to retain ultimate judgment over foreign influence due to AI's potential for similar mistakes.

Industry & market moves

Dnotitia, a Korean AI storage company, raises KRW 90 billion ($65 million USD) Series A for AI storage expansion.

Dnotitia Inc. has closed a KRW 90 billion (~$65 million USD) Series A funding round, led by Elohim Partners, to accelerate its AI storage business focused on Seahorse (a vector database) and VDPU (Vector Data Processing Unit). The VDPU is described as the world's first chip dedicated to accelerating vector data search and processing, addressing generative AI bottlenecks.

Cyber A.I. Group appoints Irving Bruckstein as CEO to drive growth and global expansion of CyberAI Sentinel 2.0.

Cyber A.I. Group has appointed Irving Bruckstein as its new CEO. Bruckstein, with over 30 years of leadership in enterprise IT and cybersecurity, will lead the company's strategic direction, oversee acquisition initiatives, and accelerate the commercialization of its proprietary CyberAI Sentinel 2.0 platform. The platform aims to provide intelligent, adaptive, and proactive protection as a low-cost alternative for small and medium-sized businesses.

Brady Corporation acquires Honeywell's Productivity Solutions and Services business for $1.4 billion.

Brady Corporation has entered a definitive agreement to acquire Honeywell’s Productivity Solutions and Services (PSS) business for $1.4 billion in an all-cash transaction. PSS provides mobile computers, barcode scanners, printing solutions, and voice guidance, generating approximately $1.1 billion in sales in 2025. This acquisition aims to expand Brady’s portfolio into mobility and scanning solutions, enabling high-volume, automated data collection and tracking for large enterprise customers in logistics, manufacturing, and retail.

Protecht acquires AI-enabled risk assessment tool VISO Trust to combat 'SaaSpocalypse'.

Sydney-based risk management software firm Protecht has acquired VISO Trust, an AI-powered risk assessment tool, to bolster its product offering against the competitive threat of AI agents and cheap alternatives. Protecht CEO Jason Phillips emphasized the need for data management tools to be resilient against AI replacement, indicating a strategic focus on AI integration to maintain competitiveness in a rapidly evolving SaaS market.

AI compute startup Parasail raises $32 million to provide elastic GPU capacity across multiple clouds.

Parasail, an AI compute startup, has secured $32 million in Series A funding to develop a global fabric that pools GPU capacity from various cloud providers and optimizes it automatically. The platform aims to solve the problem of uneven cloud GPU supply, unpredictable pricing, and complex management for AI teams, enabling developers to deploy AI endpoints in minutes.

Lua Global raises $5.8 million seed funding to build an operating system for human-AI agent collaboration.

Lua Global Inc. has secured $5.8 million in seed funding, led by Norrsken22, to develop a platform that enables businesses to build, deploy, and manage agentic AI workforces. The company aims to shift from 'workflow automation' to 'org chart,' empowering non-technical users to create and manage AI agents for various functions with built-in governance and a focus on long-term ownership rather than usage-based pricing.

Sygaldry raises $139 million to build quantum computers for AI training and inference.

Sygaldry Technologies, a quantum computing company, has raised $139 million (a $34 million seed round led by Initialized Capital and a $105 million Series A led by Breakthrough Energy Ventures). The company aims to leverage quantum computing to drastically expedite AI training and inference by combining multiple qubit types within one fault-tolerant architecture, supporting advanced AI applications at an attainable cost.

AI product & feature launches

Pervaziv AI launches Cortex 3.5, enabling multicloud AI orchestration across AWS, Azure, and Google Cloud.

Pervaziv AI has released Cortex 3.5, expanding its secure AI orchestration platform to include Microsoft Azure and Google Cloud, alongside existing AWS capabilities. This update provides enterprises with a unified AI control layer to securely access infrastructure context, orchestrate workflows, and automate operations across the three leading cloud providers. It introduces over 40 AI agents, 12 enterprise connectors, and 300+ operational capabilities.

Google Cloud announces general availability of multiple MCP servers for database and compute services.

Google Cloud has made several Model Context Protocol (MCP) servers generally available, including those for AlloyDB for PostgreSQL, BigQuery, Bigtable, Cloud Storage, Compute Engine, Managed Service for Apache Kafka, Memorystore for Redis, Memorystore for Valkey, Network Management API, Oracle Database@Google Cloud, and Pub/Sub. These MCP servers enable AI agents and applications to interact with these cloud services using conversational language, execute SQL queries, manage resources, and access data.

Aptean introduces AI platform 'AppCentral' and 10 AI agents for Business Central On-Premises customers.

Aptean has launched its AppCentral AI platform and 10 pre-built AI agents for Microsoft Dynamics 365 Business Central on-premises customers. This solution brings AI capabilities, including finance, quality, and supply chain agents, to organizations that have not migrated to the cloud, allowing partners to deploy AI in days without migration. The agents read live operational data and provide actionable recommendations in plain English, with partners able to configure up to 20 Intelligent Workflows per customer.

Infor and AWS partner to bring industry-specific AI agents to manufacturing and distribution at enterprise scale.

Infor, in collaboration with Amazon Web Services (AWS), has announced new industry-specific AI agents built natively on AWS, targeting manufacturing and distribution enterprises. These agents are designed to reason, plan, and act across complex business workflows like project management, process mining, inventory flow, and quality management. Early adopter Xpress Boats achieved a 50% reduction in expedited shipping costs and a 98% improvement in process issue diagnosis speed.

Affinda launches a no-code AI Integration Agent for document automation at scale.

Affinda Platform, an AI document processing software, has released a new AI Integration Agent. This agent allows organizations to connect AI-extracted data from documents with over 2,800 existing systems using natural language, without writing code. It enables high-stakes organizations to rapidly set up custom integrations in minutes, addressing challenges of rigid integrations or time-consuming custom API builds in regulated industries.

Superset overhauls chat UX, launches v2 early access, and introduces standalone CLI.

Superset has released a major update, featuring a top-to-bottom chat UX overhaul with a rich-text prompt editor, slash-command and file-mention chips, and a redesigned tool-call system. The v2 workspace is now in early access, offering pane layouts, a diff viewer, file editor, review tab, and browser pane. Additionally, a new Superset CLI ships as a self-contained tarball for remote server deployment, and GitHub integration is now free on all plans.

AI chatbots are habitually recommending alternative cancer treatments, sparking concern from health officials.

A new study by the Lundquist Institute found that widely used AI chatbots, including xAI’s Grok, OpenAI’s ChatGPT, Google’s Gemini, Meta’s AI, and High-Flyer’s DeepSeek, frequently recommend alternative cancer treatments to chemotherapy. Nearly half of the responses were rated 'problematic' by experts, often providing a 'false balance' between scientific and non-scientific results, which risks leading patients away from established medical treatments.

Research with immediate practical relevance

No significant new developments.