I came across this article on Hacker News calling giving a model CLI access superior to the Model Context Protocol. And it makes me feel kinda sad based on all the comments because I’ve been doing a ton of development work on MCP. But I also keep reminding myself that these guys are probably all devs with a 20x Claude Max subscription and have 6 or 7 Claude Code agents running in their ghostty on their M5 MacBook Pro and aren’t actually reviewing a single line.

Meanwhile, I’m just trying to build a preview of a new version of Siri would look like.

Like yeah, language models are pretty good at composing and executing commands. But they don’t always have that ability. And even if they do, that doesn’t mean it’s safe to do.

Claude desktop skills have the ability to run skills, but there’s no way to actually secure secrets since it’s all just a plain ZIP file

Working on this “make AI useful” project has been pretty enjoyable because it makes me feel useful since I really have to evaluate and learn how to make secure workflows using the knowledge I already have (don’t store secrets in plaintext)

It’s funny how much emerging technology starts off as horribly insecure. But following good security practices is hard, so not really. It sure isn’t fun having to take the long route of parametrizing queries when string interpolation is literally right there.

@soulfire code data confusion is a known class of security vulns and the best we can do is making a machine that fundamentally cannot differentiate between the two