Microsoft AI AI News, Models and Product Updates

Microsoft Ai news, model releases, product launches, research updates, and major announcements in one place.

Latest Articles

• Microsoft Copilot reveals secret input that allowed it to be hacked — Varonis researchers extracted a secret Microsoft 365 Copilot parameter (?autorun=1) by asking the AI to explain its own security guardrails — enabling one-click password exfiltration without user confirmation, a vulnerability Microsoft has now fully patched.
• Governed AI for Every Builder: Enterprise Controls in Snowflake CoCo — Snowflake CoCo now has per-user AI credit quotas in GA, with three more enterprise controls coming soon: MDM-pushed organization-wide policy, team-level MCP server restrictions, and tool action approval workflows — letting IT teams set guardrails without blocking builders.
• AeroCopilotBench: A Two-Tier Benchmark for Evaluating LLM Agents as Aviation Copilots in an Interactive Virtual Cockpit Environment — AeroCopilotBench evaluates LLM agents as aviation copilots on 1,200 knowledge questions and 73 emergency procedural tasks in a virtual cockpit with hard safety constraints. The best model achieves 72.6% success, and strong static knowledge consistently fails to predict safe procedural execution.
• WeSCE: A Benchmark for Measuring Security Drift in LLM-Driven Code Editing — WeSCE benchmarks how LLMs drift security posture when editing code — covering 400 real-world programs across feature addition, bug fixing, and refactoring tasks under functional-only constraints, with no explicit security requirements given to the model.
• MUSE: An Interactive Meta-Agent for Understanding and Steering LLM-powered Data Science Systems — MUSE is a meta-agent that dynamically restructures agentic data science execution traces into navigable semantic levels, surfaces suspicious steps, and scaffolds repairs without modifying the underlying agent — a between-subjects study (n=15) showed improved task efficiency and user confidence (UIST 2026).
• Designing a Persistent Knowledge Layer That Refuses to Guess — A detailed practitioner blueprint shows how to build a persistent knowledge layer on top of RAG so AI systems accumulate understanding across queries instead of discarding reasoning after each response — using Azure AI Search, Cosmos DB, Microsoft Foundry, and FastAPI on a property-insurance corpus.
• SpaceX officially closes its Cursor acquisition — SpaceX officially closed its $60 billion acquisition of AI coding tool Cursor, giving the startup access to SpaceX's GPU fleet — currently rented by Anthropic and Google — and completing Elon Musk's vertical integration of AI infrastructure alongside xAI, acquired earlier this year.
• Why AI Governance Frameworks Are Hard to Adopt: A Role-Based Stress Test of the NIST AI RMF — A role-based stress test of the NIST AI Risk Management Framework in consumer lending finds the framework fits bounded ML models cleanly but poorly for workflow-embedded LLM copilots — structural fit, not adoption intent, is the primary barrier to governance value.
• Microsoft’s Clippy-like Mico character is no longer the face of Copilot — Microsoft is retiring Mico, the animated yellow blob that gave Copilot's voice mode a face, less than a year after launch — the character moves to the Learn Live education platform as Microsoft consolidates Copilot and Microsoft 365 Copilot into one product.
• Microsoft is combining its Copilot apps ahead of a ‘super app’ — Microsoft is merging consumer Copilot and Microsoft 365 Copilot into a single app starting mid-August, retiring Podcasts and Deep Research on August 18th as groundwork for a broader Copilot super app planned for later in 2026.