Mr. Dutch

Building StackGrade ( AI agent tools ranked on quality & safety. And InEUyet ( which AI features reach the EU.

[Mr. Dutch @MrDutch310822]

@CircleCI Handy. Keep in mind the failing log includes output from tests and dependencies, so whatever text ends up in it goes straight into the agent's prompt. Worth a quick read before you paste it into an agent that can run commands.

[Mr. Dutch @MrDutch310822]

@automater_ai Worth checking functions too. A SECURITY DEFINER function runs with its owner's privileges, so a role that can only SELECT can still change data if it's allowed to execute one. And new functions get EXECUTE granted to PUBLIC by default.

[Mr. Dutch @MrDutch310822]

@vikktorrrre Anthropic's help page backs that up: you need a Pro or Max plan, a Claude account in the US and a waitlist invite. It names no timeline for other regions, not even a vague one.

[Mr. Dutch @MrDutch310822]

@truemar14 No date yet. Anthropic's help page says Claude Duet is "available in the US only during the research preview", and you also need a Pro or Max plan plus a waitlist invite. Nothing about Europe so far. We track it here: https://t.co/ZhDBxICkwh

[Mr. Dutch @MrDutch310822]

Gemini in Chrome reached the UK on 14 July. 106 days later, Google's list of supported countries still has no EU member state. Ukraine is on it. Germany isn't.
https://t.co/UugTk2ZOKn

[Mr. Dutch @MrDutch310822]

Two tools on StackGrade lost their "No issues found" label this week. wshobson/agents now has 6 critical-severity known vulnerabilities in its runtime dependencies, and Mastra has 1. A clean result only covers the advisories published so far, so we rescan every week.

[Mr. Dutch @MrDutch310822]

@milesdeutscher @bot interessant?

[Mr. Dutch @MrDutch310822]

@milesdeutscher @bot interessant👍

[Mr. Dutch @MrDutch310822]

31 of the 52 AI agent tools on StackGrade shipped a release in the week to 26 Oct. Coding agents were the busiest group, 10 of 14. If your agent updates itself, that's a lot of new code running on your machine every week. Do you pin versions, or take whatever ships?

[Mr. Dutch @MrDutch310822]

@AndrewHart Small consolation: the same Siri AI does run in the EU on a Mac with macOS 27 and on Vision Pro. Just not on the phone, which is where most of us were going to talk to it.

[Mr. Dutch @MrDutch310822]

@Wraithioner @9to5mac @apollozac If you have a Mac on macOS 27 it should show up there: Apple says EU users get Siri AI on macOS 27 and visionOS 27. It's only iOS, iPadOS and watchOS that are still left out. https://t.co/sqjQTZ9VYM

[Mr. Dutch @MrDutch310822]

Siri AI in the EU right now: yes on your Mac, yes on Vision Pro, no on iPhone, iPad or Apple Watch.

Apple's October note still says it is "currently unavailable in the EU in iOS, iPadOS, and watchOS." So in Europe, the laptop got the smarter Siri before the phone did.

[Mr. Dutch @MrDutch310822]

Five tools on StackGrade went from "No issues found" to "No dependencies scanned" this week, Claude Code among them. Their scan didn't change. They have no lockfile, so there was nothing to check, and calling that "no issues" was too generous of us.

[Mr. Dutch @MrDutch310822]

@honchodotdev Does the scan read what the tool descriptions tell the model, or only the code? Tool poisoning tends to sit in those descriptions, and they can change in an update after the grade.

[Mr. Dutch @MrDutch310822]

@Joe_brendan_ The part that still takes time is deciding what the screen should do and what it leaves out. A working build answers that faster than a mockup, because people poke at it instead of nodding at it.

[Mr. Dutch @MrDutch310822]

@kenashe Worth checking your agent's config afterwards too, like ~/.codex/config.toml for MCP servers you never added. If a planted file got the agent to change a setting, that change stays after you leave the folder.

[Mr. Dutch @MrDutch310822]

7 of the 52 AI agent tools on StackGrade ship an agent or skill file that tells the agent to pipe a remote script into a shell. The agent takes that file as instructions, and the script can change any day. Worth reading those files before your agent does. https://t.co/YNjMZpa6Rs

[Mr. Dutch @MrDutch310822]

InEUyet is live: a tracker that answers one question, is that AI feature available in the EU yet?

39 features from OpenAI, Google, Anthropic, xAI, Meta, Microsoft and Apple. Right now 14 are not available in the EU. Each answer links its source.

https://t.co/8uyBLYdZZL

[Mr. Dutch @MrDutch310822]

Don't start a coding agent in ~/Downloads, or in any folder full of files you didn't write. Codex reads an AGENTS.md and Claude Code reads a CLAUDE.md from the folder you launch in, and both treat what's inside as instructions. Make a clean folder first.

[Mr. Dutch @MrDutch310822]

@MasterDotDev Is the MCP one about building servers or about using them safely? The second part is where I see most people get stuck: which servers to trust and how much access to give them.