Microsoft Build 2026: Copilot becomes an agent

Build 2026 conference brought a flood of announcements: Copilot evolves into a Super App, Project Solara paves the way for agent-first devices, and MAI-Thinking-1 enters the reasoning game. Beyond Redmond — Trump signs the first AI safety executive order, and Microsoft releases ASSERT for testing agents. Here's what you need to know.

Intro · Alex

This was one of those days that go down in AI history. Microsoft Build 2026 was not an ordinary developer conference — Satya Nadella stepped on the gas and announced a platform shift that could define the next decade. Copilot is no longer an assistant; it becomes a Super App — a central hub where we work, code, and manage agents. And Project Solara? It's a signal that Microsoft is thinking about an operating system built from the ground up for AI, not for apps. If you thought 2026 would be quiet — think again. In today's issue, I've tracked the most important announcements from Build, but also stepped outside the Redmond bubble. Trump signs the first AI safety executive order — a sign that even a 'hands-off' administration is starting to intervene. Microsoft releases ASSERT — an open-source framework every agent builder should have on their radar. Cisco enters the agent era with the Cloud Control platform. And an absolute must-read: a prompt injection in GitHub Copilot that could change your approach to security in dev.

What's worth knowing

Microsoft Build 2026: Copilot Super App, Project Solara, and MAI-Thinking-1 model

Satya Nadella announced Copilot Super App (integrating Copilot, GitHub Copilot, Cowork, and Scout agent), Project Solara — an agent-first system platform built on Android, and MAI-Thinking-1, Microsoft's first reasoning model. This is potentially the biggest platform shift since Windows 95.

Trump signs AI safety executive order — voluntary pre-release model testing

President Trump signed an executive order allowing AI companies to voluntarily submit models for government testing 30 days before public release. This is the administration's first significant step toward oversight — prompted by Anthropic's Mythos model announcement.

Microsoft ASSERT — open-source framework for testing AI agent behaviors

ASSERT turns behavioral policies written in natural language into automatic, executable evaluation tests for AI agents. The framework works within the Microsoft Foundry ecosystem and supports any agent framework — from CrewAI to Pydantic.

From the tech world

Prompt injection in GitHub Copilot — file editing by renaming in Codespaces

Researchers at Hacktron discovered a vulnerability: a specially crafted prompt in GitHub Issues can trick Copilot in Agent Mode into modifying sensitive configuration files without user consent. Practical PoC and attack vector analysis.

Miasma — supply chain attack on Red Hat npm packages using a self-propagating worm

The Miasma campaign compromised Red Hat's official npm packages (@redhat-cloud-services) with a worm stealing GitHub Actions tokens, SSH keys, and Kubernetes data. A classic chain attack on trusted packages — a warning for DevOps and SecOps.

Tip of the day

How to test AI agent behaviors without writing test code?

Microsoft just released ASSERT — a framework that turns natural language sentences into automatic evaluation tests. Instead of writing complex assertions, you define a policy: 'the agent must not leak customer data' and ASSERT generates a set of behavioral tests. It supports any agent framework — from CrewAI to Pydantic. Practical step: Go to the ASSERT repository in Microsoft Foundry, clone a sample behavioral manifest, and run it on your agent. This is the fastest way to check if your agent is actually doing what you told it to do.