AI labs ship new models, infrastructure hardens around agents, and financial markets get their first MCP data feeds.
The week of April 10th to April 17th saw Anthropic ship a new flagship model with live security controls and enter the design tool market, OpenAI push AI into life sciences research, and Cloudflare lay down the networking foundation that production-grade agents have been missing.
TL;DR
Anthropic released Claude Opus 4.7 with stronger coding and vision capabilities and new automated cybersecurity safeguards, and launched Claude Design, a research preview product for creating prototypes, slides, and visuals through conversation. OpenAI upgraded its Agents SDK with memory and sandbox tools, expanded Codex into a general workspace, and released GPT-Rosalind, a gated life sciences model that outperforms GPT-5.4 on biology research tasks. Claude Code Routines launched for scheduled and event-driven coding automation on Anthropic’s cloud. Microsoft OneLake MCP reached general availability inside Microsoft Fabric.
Cloudflare launched an enterprise-grade hosted MCP platform alongside Cloudflare Mesh, a private networking layer that gives each agent a distinct identity. Google released agentic developer tools for Android. Nymbus launched one of the first MCP servers built for core banking, and Insig AI put 60 central banks’ data behind an MCP server. IBM launched IBM Autonomous Security to fight agentic attacks at machine speed. GitHub CLI added the gh skill command for versioned agent skill management. Red Hat published MCP security guidance for OpenShift. Figma released a plain-English MCP guide. Zenskar raised $15M Series A for AI-native B2B billing automation with MCP integration.
Major Updates of the Week
Claude Opus 4.7 With Built-In Cybersecurity Safeguards
Anthropic released Claude Opus 4.7 on April 16th across all Claude products, the API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry at the same price as Opus 4.6. The model brings gains in agentic coding, vision resolution, and long-horizon task management, plus a new “xhigh” reasoning effort level. It is the first Anthropic model to ship with automated safeguards that detect and block high-risk cybersecurity uses in real time. What Anthropic learns from this rollout will directly shape how it approaches a broader release of Mythos-class models.
Claude Design: AI-Powered Prototypes and Visuals for Every Team
Anthropic launched Claude Design on April 17th as a research preview for Pro, Max, Team, and Enterprise subscribers. Users describe what they need and Claude builds a working design, refined through conversation or direct edits. Claude reads a team’s codebase during onboarding to apply a shared design system automatically across every project. Outputs export as PDF, PPTX, HTML, or go directly to Canva or Claude Code. Early users at Brilliant and Datadog compressed multi-day design cycles into a single session.
OpenAI Updates
| Vendor / Product | Key Action / Feature | Significance |
| Agents SDK Evolution | Added configurable memory, sandbox-aware orchestration, and Codex-like filesystem tools for long-running agents | Reduces infrastructure overhead so developers focus on domain logic rather than core plumbing |
| Codex Expansion | Codex now includes an in-app browser, computer use for macOS apps, and project-free chat threads | Shifts Codex from a coding tool to a general workspace for getting work done with AI |
| GPT-Rosalind | First life sciences model, outperforming GPT-5.4 on 6 of 11 biology tasks. Connects to 50+ scientific databases via a free Codex plugin. Gated to qualified US enterprise customers | Moves AI from general assistant to genuine research collaborator for drug discovery and genomics |
Microsoft OneLake MCP Is Now Generally Available
Microsoft’s OneLake MCP server reached general availability inside Microsoft Fabric this week. Agents can now browse, read, write, and query data across the full OneLake data estate through conversation, using a team’s existing Azure identity and permissions. No new roles or access are required. Admins, engineers, and analysts can all interact with their data through natural language without writing a query or navigating the portal.
Claude Code Routines: Automate Coding Tasks Around the Clock
Anthropic launched Claude Code Routines, letting developers configure a prompt once and run it on a schedule, via GitHub event, or API call. Unlike previous scheduling tools, routines run on Anthropic’s cloud infrastructure, so they continue working even when the local machine is off. Pro users get 5 runs per day, Max users get 15, and Team and Enterprise users get up to 25. Claude Code effectively becomes a background service rather than a tool that needs manual activation.
Other Updates
| Vendor / Product | Key Action / Feature | Significance |
| Cloudflare Mesh | Private networking layer that assigns each AI agent a distinct identity and enforces granular access policies across cloud infrastructure | Removes the choice between slow VPNs and unsafe open exposure, giving agents secure access from day one |
| IBM Autonomous Security | Multi-agent security service that detects, enforces policy, and remediates threats at machine speed across an enterprise’s full security stack | Responds to attackers already using AI to accelerate every phase of the attack lifecycle |
| Nymbus MCP Server | One of the first MCP servers built for core banking, with 19 tools for customer lookup, account management, money movement, and card controls | Gives US banks and credit unions a single, governed interface for AI-assisted front-office workflows |
| Insig AI Central Bank MCP | MCP server connecting LLMs to 60 central banks’ communications spanning 10 years and 15,000 documents | Collapses days of analyst research into real-time, on-demand signals for portfolio managers and traders |
| GitHub CLI: gh skill | New gh skill command installs, versions, and pins agent skills with the same guarantees as a package manager | Treats agent skills as auditable, supply-chain-safe artifacts rather than loosely managed files |
| Red Hat OpenShift MCP | Published guidance on containerizing MCP servers securely within OpenShift environments | Gives enterprise teams a clear security and deployment model for running MCP at scale |
| Figma MCP Guide | Plain-English explainer of what MCP is and how it connects to design workflows | Lowers the barrier for non-technical teams to understand and adopt MCP in their tools |
| Zenskar $15M Series A | AI-native billing platform raises $15M led by Susquehanna and Bessemer to expand its agent marketplace and MCP integrations | Finance teams can deploy agents across the full billing cycle without any engineering support |
Cloudflare Enterprise MCP: Managed Hosting for Production Agent Infrastructure
Cloudflare launched a separate enterprise MCP hosting platform this week, distinct from Cloudflare Mesh. It provides a managed, remote environment for MCP servers with built-in access controls, authentication, and rate limiting. Organizations can expose governed MCP endpoints to their agents without running any local infrastructure. Together with Cloudflare Mesh, teams now have a full stack: a secure private network for agents to operate within, and a managed layer for exposing the tools those agents need.
Google Brings Agentic AI Tools to Android Studio
Google released new agentic AI tools for Android developers this week, targeting a three times faster app build cycle using any compatible AI agent. The tools connect agents directly to Android Studio, letting them understand project structure, generate code, and run builds without manual setup. This positions Android Studio as an agent-ready environment and signals that Google is treating coding agents as a core part of its developer platform rather than an optional add-on.
My Thoughts: The Infrastructure Race Behind AI Agents
The biggest story this week is not any single product but the fact that model capability and agent security moved forward together. Claude Opus 4.7, Cloudflare Mesh, and IBM Autonomous Security are not independent releases. They reflect the same reality: as agents gain more access, the infrastructure around them has to harden in parallel.
The financial sector moves are the quieter signal worth watching. Nymbus and Insig AI putting MCP inside core banking and central bank data feeds means regulated industries are past the pilot phase. When MCP reaches the ledger, it carries real accountability.
