AI Week 21-28 April 2026: Google's $40B Anthropic Bet, Copilot Goes Agentic, and a New Cybersecurity Front

Four stories defined the AI market this week: Google's record investment in Anthropic, Microsoft Copilot Agent Mode going GA across Office, a heavyweight AI cybersecurity coalition, and a deeper AWS–Anthropic partnership. Here is a fast, practical read for executives.

1. Google bets up to $40B on Anthropic

On April 24 Google announced an immediate $10B cash investment in Anthropic at a $350B valuation, with up to another $30B contingent on performance milestones. On top of that, Google Cloud will provide 5 gigawatts of compute capacity over the next five years. It is one of the largest AI deals to date and cements Anthropic among the leading frontier model providers.

What it means for organisations: Access to frontier AI is deepening but is also concentrating in a small number of providers. Build your AI stack with model and cloud portability in mind — single-vendor lock-in is now an explicit board-level risk.

2. Microsoft Copilot Agent Mode is GA in Word, Excel and PowerPoint

On April 22 Microsoft made Copilot Agent Mode generally available in Word, Excel and PowerPoint on Windows, Mac and the web. Copilot can now execute multi-step, app-native actions inside documents, workbooks and decks — taking users from first draft to final output while keeping a human in the loop. Microsoft reports 67% growth in Excel engagement and 65% higher user satisfaction during preview.

CribConnects take: This shifts Copilot from "smart assistant" to "executing colleague" in everyday Office work. Now is the time to redesign standard reporting, analysis and presentation workflows — otherwise the value gets stuck at the prompt level.

3. Anthropic launches Project Glasswing for defensive cybersecurity

On April 27 Anthropic introduced Project Glasswing, a coalition including AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, NVIDIA, Palo Alto Networks and the Linux Foundation. Members get early access to Claude Mythos Preview for vulnerability detection, black-box testing and penetration testing. Anthropic is putting up to $100M in usage credits and $4M in donations behind the effort, with the first public findings expected within 90 days.

What to do: Security teams should plan for a world where both attackers and defenders run frontier models. Threat models, detection engineering and red-teaming deserve a refresh this quarter.

4. AWS and Anthropic go deeper on the chip layer

In the April 27 AWS Weekly Roundup, AWS confirmed that Anthropic now trains its most advanced foundation models on AWS Trainium and Graviton, co-engineered at the silicon level with Annapurna Labs. Claude Cowork is available in Amazon Bedrock, with a unified Claude Platform on AWS coming soon. Combined with the Google deal, the picture is clear: Anthropic is positioning itself as a multi-cloud frontier player.

---

The week in one line

Infrastructure, distribution and the application layer all moved this week. For European organisations the three big questions are: how do you ensure vendor diversity, how do you capture real value from agentic Office, and how do you professionalise AI security? CribConnects helps you turn this week's signals into a concrete roadmap.

Want to talk it through? Book an AI Scan at cribconnects.com and turn this week's news into action.