Back to updates

Commandline - 1

2026 (12)

October (7)

commandline: Copilot to autopilot: Why no single safeguard is enough

Microsoft Command Line Blog

2026-10-13

Microsoft says it built a layered control model for an AI agent to handle production network incidents autonomously, but only within defined identity, authorization, tooling, and human-approval boundaries. The article argues that no single safeguard is sufficient for production use.

commandline: Introducing Microsoft-Decision-1, our model for fast decision-making

Microsoft Command Line Blog

2026-10-10

Microsoft introduced Microsoft-Decision-1, a decision-scoring model for structured decision tasks. It is described as optimized for low latency and high quality, and positioned as outperforming both LLMs and other decision models.

commandline: Bringing local models and sandboxed tools to Windows and GitHub Copilot

Microsoft Command Line Blog

2026-10-07

GitHub Copilot is adding support for deciding when to use on-device models versus cloud models on Windows. The update also mentions sandboxed tools, but no further details are provided in the excerpt.

commandline: Beyond violation rates: Are your AI evaluations measuring the right things?

Microsoft Command Line Blog

2026-10-05

The article argues that structured evaluation design can make AI evaluations more comprehensive and easier to interpret. It emphasizes measuring a wider range of behaviors and producing failures that are traceable and actionable.

commandline: Building an agentic system and a learning loop around customer IP

Microsoft Command Line Blog

2026-10-05

Microsoft and Kantar built an agentic system around Kantar’s 30 years of ad metrics and frameworks to evaluate and help optimize advertisements. The system is described as part of a learning loop for customer IP.

commandline: Azure Content Understanding: Building on lessons from a decade of Document Intelligence

Microsoft Command Line Blog

2026-10-02

Microsoft reflects on a decade of document understanding work and how those lessons inform Azure Content Understanding. The piece emphasizes limits of LLMs in document reading and the importance of handling edge cases with confidence, grounding, and reasoning.

commandline: LLM-as-translator pattern: Adding natural language understanding to an existing system without rewriting it

Microsoft Command Line Blog

2026-10-02

The article describes using an LLM as a translation layer to add natural-language understanding to an existing system without rewriting it. The approach modernizes legacy implementations by inserting an abstraction layer instead of replacing core logic.

September (2)

commandline: Introducing run-assert-eval: Find the risk, fix it, prove it

Microsoft Command Line Blog

2026-09-24

Microsoft introduced run-assert-eval, a skill for runtime risk discovery in agents. It identifies relevant risks, measures failure rates, generates a runtime policy from the results, and reruns evaluation to verify the fix.

commandline: PyRIT: Democratizing AI red teaming through open-source tooling

Microsoft Command Line Blog

2026-09-10

PyRIT is presented as an open-source Python tool for AI red teaming that lowers the barrier for subject matter experts to test generative AI systems. The article describes its evolution toward making red teaming more accessible across many domains.

August (3)

commandline: Designing product simplicity for the AI era

Microsoft Command Line Blog

2026-08-31

The article asks whether shared design components are sufficient for a seamless product experience, or whether designers need to prioritize simplicity over complexity to create consistency. It frames this as a product design question for the AI era.

commandline: Your agent’s guardrails have a bypass

Microsoft Command Line Blog

2026-08-27

Microsoft introduced Agent Hooks, an open, framework-neutral governance contract for agent systems. It includes conformance testing and is intended to make enforcement fail closed instead of failing open when controls crash.

commandline: One requirement, many failure paths: Evaluate, control, and optimize with ASSERT and ACS

Microsoft Command Line Blog

2026-08-24

The article says a safety requirement can have a complex runtime failure surface even when it seems simple. It highlights ASSERT for finding coverage bugs and ACS for fixing the policy.