August 23, 2026
Today's top tech stories, deduped across the newsletters I read and briefly summarized. Click a source to open the original article.
AI
Model Context Protocol publishes new roadmap
The team behind the Model Context Protocol (MCP) has published an updated roadmap for the standard that lets AI models connect to tools and data sources. The plan outlines upcoming features and priorities for the ecosystem.
Linus Torvalds: AI helped with a 'debugging session from hell'
Linux creator Linus Torvalds describes how an AI did much of the grunt work in an unusually difficult kernel debugging session, though he notes the tool could also be flatly wrong with total confidence.
The key skill for coding agents: instructing and verifying, not just reading code
Simon Willison argues the most important skill for using coding agents productively is confidently instructing them and then verifying the changes are correct, not necessarily reviewing every line they write.
A week of using Codex more than Claude
A developer shares takeaways from a week spent using OpenAI's Codex more than Claude for coding tasks, comparing the strengths and weaknesses of the two coding agents.
Why your local LLM feels dumber than it is
A Level1Techs forum post explains why locally run language models often underperform expectations, pointing to factors like quantization, context window, and prompt setup as more important than raw model size.
Security
Hackers infect Android car head units with proxy botnet malware
A supply chain attack against Android-based car head units uses a legitimate device-update app to spread malware that enlists compromised devices into a proxy botnet or uses them for ad fraud.
Named pipes under attack: securing Windows IPC
Windows named pipes provide fast interprocess communication, but weak access controls can expose privileged services to untrusted code. ThreatLocker explains how endpoint verification, command authorization, and narrowly scoped privileges can help secure the communication.
Dev
Simon Willison's llm tool updated after OpenAI library breakage
After OpenAI's Python library dropped its use of httpx, fresh installs of the llm command-line tool stopped working. Simon Willison shipped a quick fix in version 0.32.1, followed by 0.33, which upgrades to OpenAI library 3.x and switches the HTTP client to httpx2.
llm-openrouter 0.7 shows LLM reasoning traces
The llm-openrouter plugin is now compatible with llm 0.32 and can display the reasoning traces of language models available through OpenRouter.
'Stop making TUIs': build real interfaces with coding agents instead
Thomas Ptacek argues developers should build real, native user interfaces even for the smallest personal tools, since coding agents have made it far cheaper to put together a usable GUI than to fall back on text-based interfaces (TUIs).
Rust Glancer: new Rust LSP uses 100x less RAM
The new Rust Glancer project promises a Rust language server that uses roughly 100 times less RAM than existing alternatives, which could make large Rust codebases far snappier to work with in editors.
IT
Apple deprecates hdiutil in macOS 27 'Golden Gate'
Apple has marked the hdiutil command-line tool, used to manage disk images on Mac, as deprecated in the upcoming macOS 27 'Golden Gate,' worrying developers who rely on it in scripts and build pipelines.