Browse Artificial Intelligence News (437)

Microsoft Threat Intelligence details CaptiveCrunch, a Storm-2945 (Midnight Blizzard) campaign abusing captive portals to manipulate traffic, deliver malware, and steal credentials and tokens. The post breaks down the CornFlake and ChocoShell toolchain, device code phishing against Microsoft Entra ID, and provides Defender/Sentinel hunting queries plus practical mitigations for travelers and enterprises.

Allison announces the deprecation of the Gemini 2.5 Pro and Gemini 3 Flash models across GitHub Copilot experiences, and explains what Copilot Enterprise admins may need to change in model policies so users can select supported alternatives in VS Code and on github.com.

Allison announces a public preview for GitHub Enterprise that lets Copilot admins target AI model policies to enterprise teams, so different groups can be granted additional models beyond an enterprise-wide baseline.

Samantha Kubota introduces Echoverse, a set of high-fidelity synthetic “worlds” for training computer-use agents, and explains why depth (stateful workflows and database-grounded grading) matters more than sheer environment count. The post shares results from supervised fine-tuning and reinforcement learning, plus links to released code and datasets.

Amaury Levé introduces an open-source polyglot unit-test generation agent that learns a repo’s conventions, writes tests, and validates they build and run in the project’s normal test workflow. The post explains the agent’s research/plan/execute/check loop, shows benchmark results versus stock Copilot, and includes setup commands for Copilot CLI and VS Code.

Kristen Womack’s July 2026 roundup for Azure Developer CLI (azd) summarizes releases 1.27.0 through 1.29.0, including new extension install options, a full tool uninstall workflow, Azure AI Foundry resource modeling in azure.yaml, container deployments to Azure App Service, and CI-friendly non-interactive behavior plus a breaking flag rename.

diptiborkar summarizes why Microsoft Fabric and OneLake were recognized as a Leader in Forrester’s 2026 Data Lakehouses report, focusing on a unified lakehouse foundation for analytics and AI, open table formats, cross-cloud interoperability, and built-in governance and security across engines like Spark, SQL, KQL, and Power BI.

Allison announces the retirement of GitHub Models as of July 30, 2026, including the shutdown of the playground, model catalog, inference API, and BYOK support. The post also points teams to Microsoft Foundry for model access and GitHub Copilot for building AI-powered workflows on GitHub.

Samantha Kubota explains how Microsoft is pushing better AI performance per token and per dollar by co-optimizing models with evaluation harnesses and RL environments, sharing recent “Flash” model releases and benchmark results that reduce GPU costs while maintaining quality across products.

Cassidy Williams shows how she used the GitHub Copilot app to modernize an old React codebase by splitting work into stacked sessions and stacked pull requests, keeping changes reviewable while iterating on plans, branching strategy, and dependency upgrades.

Shawn Henry and Tamir Dresher show how to run a multi-agent “Squad” team on top of the GitHub Copilot CLI/SDK and expose it as a first-class Microsoft Agent Framework (MAF) AIAgent in a .NET app, including DI wiring, session configuration, and OpenTelemetry-based observability.

Alym Rayani rounds up July 2026 updates across Microsoft Security, focusing on securing AI and agent environments. The post highlights Project Perception’s agentic SecOps approach, new Microsoft Defender protections (including prompt-injection defenses and cloud agent runtime coverage), Entra identity updates like passkeys-by-default, and Purview controls for data protection and investigations.

Allison summarizes the July 2026 updates for GitHub Copilot in Visual Studio, including a new preview agent built on the Copilot SDK, built-in .NET and Azure skills, selected-code review actions, and organization-wide custom instructions for consistent Copilot behavior across teams.

Allison announces a new GitHub Copilot enterprise managed setting that lets organizations restrict which managed devices can host remotely controlled Copilot sessions, including options to require SSO, disable remote control, or allow it without restrictions.

Allison summarizes the July 2026 GitHub Copilot updates in Visual Studio Code (v1.127–v1.131), focusing on the Agents window workflow, multi-chat sessions, model and chat controls (including Copilot Vision and BYOK), plus editor, terminal, browser, and accessibility improvements.

Allison announces general availability for Copilot code review support for agent skills and MCP server connections, enabling teams to bring internal standards and external context into reviews. The update includes attribution on generated comments and outlines where to configure MCP servers and store tokens in repository settings.

murray-kp rounds up the July 2026 Microsoft Fabric updates across CI/CD and Git integration, OneLake governance, Spark runtime and diagnostics, lakehouse and warehouse features, and Real-Time Intelligence improvements including Eventstream security and new connectors.

Rachel Kang introduces a new “thinking effort” control in Visual Studio 18.9 Insiders 2 that lets developers tune how much reasoning supported AI models use per request. The post explains the available levels, where to configure them in the model management UI, and how higher reasoning settings affect usage credits and response depth.

Allison announces a new GitHub Copilot policy that automatically enables newly generally available models by default for Copilot Business and Copilot Enterprise, with a single opt-out control for organizations that need stricter governance.

Faith Xu shares early production results for MAI-Code-1-Flash in GitHub Copilot Chat in VS Code, focusing on the trade-off between coding quality and token efficiency. The post explains the metrics used (acceptance, code/commit survival, tokens per turn, turns per commit) and what the team learned from aggregate usage data and A/B “Auto” flights.

Allison announces expanded GitHub Copilot app reporting in the Copilot usage metrics API, adding user attribution and rollups across features, models, and languages, plus token and code activity breakdowns to help enterprises compare Copilot surfaces consistently.

Jeff Handley announces MCP C# SDK v2.0, which implements the 2026-07-28 MCP specification revision. The post explains the new stateless-first HTTP transport, standardized routing headers, and Multi Round-Trip Requests (MRTR) for interactive tools, plus what stays backward compatible and what changes for Tasks and other extensions.

Allison announces that Grok 4.5 (xAI’s reasoning model) is rolling out in GitHub Copilot, highlighting its large context window, multimodal inputs, and suitability for agentic, multi-step coding workflows. The post also covers where you can select the model and how admins enable it for Business and Enterprise.

Mark Downie introduces the Visual Studio July update, centered on a new Agent (Preview) in Copilot Chat built on the GitHub Copilot SDK. The post also covers built-in .NET and Azure skills, org-level Copilot custom instructions, attaching Git branches as chat context, and improved MSVC toolset discovery for reproducible C++ builds.

Sergey Menshykh explains how .NET agents can discover and load Agent Skills directly from a Model Context Protocol (MCP) server using the Microsoft Agent Framework, so teams can publish skills centrally and have agents pull them on demand without redeploying.

Allison announces updates to GitHub Copilot for JetBrains, including configurable OpenTelemetry export for agent workflows, tighter controls for model behavior (token limits, enabling/disabling built-in models), and support for MCP servers and custom agents in Claude agent flows, plus Copilot CLI and UX improvements.

Allison announces a new dedicated policy for the GitHub Copilot app, letting enterprises and organizations control access independently from Copilot CLI. The update keeps existing guardrails in place by routing changes through isolated workspaces and pull requests, so reviews, checks, and audit history still apply.

Burke Holland shares a practical GitHub Copilot “harness” workflow for building features end-to-end: pick a Copilot surface, enable more autonomy safely, prototype and plan in the same session, implement with Autopilot, then iterate and run a Rubber Duck review to catch blind spots.

AmirJafari announces updates to Fabric data agents in Microsoft Foundry, focusing on simpler MCP-based connectivity, OneLake Catalog discovery, support for attaching multiple data agents to a single Foundry agent, and improved production troubleshooting via Foundry Observability traces and logs.

Satya Nadella announces MAI-Cyber-1-Flash, a cybersecurity-focused model aimed at finding difficult vulnerabilities in large codebases, and describes how it pairs with MDASH to reduce cost while maintaining strong results. He also introduces Project Perception, an agentic security offering built around real-world signals and security workflows.

Allison announces that enterprise managed settings now apply to the GitHub Copilot app and Copilot cloud agent, letting enterprise owners enforce a single managed-settings.json policy across more Copilot clients, including controls for plugins, marketplaces, and approval-prompt bypass behavior.

Yuliia Kovalova and Jan Krivanek introduce the MSBuild Binlog Analyzer for VS Code, a preview extension that uses GitHub Copilot Chat plus an MCP server to explain build failures, pinpoint performance bottlenecks, compare builds, and spot regressions directly from MSBuild .binlog files.

mayurjain explains how the OpenAI Assistants API retirement on August 26, 2026 affects programmatic integrations with Microsoft Fabric Data Agents, and why Microsoft recommends moving to the Data Agent Model Context Protocol (MCP) endpoint instead of rebuilding directly on OpenAI’s Responses API.

sbaynes introduces MAI-Cyber-1-Flash, Microsoft’s compact cyber-focused model used inside MDASH to handle most security tasks at lower token cost, while reserving larger models for harder cases. The post also outlines Project Perception’s agentic security workflows and the safety, governance, and enterprise controls built into deployment.

Ram Shankar Siva Kumar announces Microsoft’s External Red Team Alliance (EXTRA), a global effort to expand AI red teaming beyond internal teams by funding academic research and building an external network of specialists to test frontier AI systems across languages, regions, and domains.

Christopher Harrison introduces the GitHub Copilot app and shows how it supports real-world development workflows with project-scoped agent sessions, Quick Chat for parallel threads, interactive canvases for visual iteration, and Agent Merge to help move pull requests through CI and review.

Shireesh Thota summarizes what customers say they value most in Microsoft’s database portfolio, and ties that feedback to recent and upcoming investments across SQL Server, Azure SQL Database, Azure Database for PostgreSQL, and Azure Cosmos DB—especially around reliability, scalability, and AI-ready capabilities like vector search and RAG.

Allison announces that Anthropic’s Claude Opus 5 is now selectable in GitHub Copilot, highlighting its strengths on long-running, multi-step coding tasks and agentic workflows, plus where it’s available, how admins enable it for Business/Enterprise, and how usage-based billing applies.

Allison announces general availability of Copilot cloud agent for Linear, letting teams assign Linear issues to Copilot to produce draft pull requests, run work in an ephemeral environment backed by GitHub Actions, and report progress back into Linear with configurable models, branches, and custom agents.

arindamc explains how to build a real-time “row-to-intelligent-action” pipeline in Microsoft Fabric that captures shopping-cart signals via CDC, reshapes them into analytics-ready rows, and enriches each micro-batch with AI in a Spark Structured Streaming notebook to drive timely actions.