News on technology, AI, Claude, .NET and QA, summarized in my own words and always linking the original story. For long technical write-ups, see the Articles.
Wednesday to Friday was about keeping agents under control: Claude can now limit where a key works, Cloudflare split out traffic from personal assistants, Google measured what a TODO comment costs a coding agent, and Copilot got a path for moving .NET apps to Windows Arm64.
With the GitHub Universe news digested, Thursday and Friday brought more practical things: bigger limits and a cost API for Claude, another free Blazor toolkit, lockfiles for GitHub actions, an open 1-trillion-parameter model and a QA agent that tries to spend less AI on every run.
Kimi's new model reaches the API with a 1-million-token context window and coding scores close to the leading closed models.
The September update, published on 10/22, brings Bring Your Own Model in preview, pull request exploration through the Git agent and a shortcut to update dependencies with security flaws.
After MCP, it is now the turn of skills: the SKILL.md-based format gets a neutral home, with maintainers from several companies.
In research preview for Pro and Max subscribers in the US, Claude Duet runs on its own virtual machine, uses connected apps and hands finished work back for review.
Anthropic released the first new API version since 2023-06-01, and the new major versions of the SDKs and the ant CLI already use it by default.
Microsoft announced the Type Server Protocol, an open standard that separates the Python extension from the type checker, and a partnership with Meta to use Pyrefly instead of Pyright.
Microsoft introduced an open source framework, based on Model-View-Update, where the UI of WinUI apps is described in C# code and rendered with real WinUI controls.
Copilot Autofix in GitHub Advanced Security for Azure DevOps now runs on the Copilot coding agent, proposes fixes for vulnerable dependencies and accepts alerts from third-party tools.
Richard Socher's startup released Flint, a 40-billion-parameter MoE model (2.5 billion active) under the Apache 2.0 license, which the company says was largely built by an automated research system.
On October 21 the PHP Foundation released versions 8.5.12, 8.4.27, 8.3.36 and 8.2.35, which fix two remote code execution vulnerabilities rated CVSS 9.2.
Forged certificates could impersonate any site; Firefox, Chrome, Apple and Microsoft have already blocked the affected CAs, but a lot of non-browser software still needs updating.
Instead of locking the whole repository on every write, Spokes now serializes per reference and gains virtual replicas backed by Azure Blob Storage.
The model makes predictions such as churn, demand and fraud without training for each task, and ships with an open library and weights on Hugging Face.
A Stacklok survey of 520 technical leaders shows that integrations, MCP servers and agent workflows weigh more on lock-in than the model itself.
Physicist Christopher Stubbs has joined the Long-Term Benefit Trust, the independent group that helps Anthropic balance commercial growth with its public benefit purpose.