There is no shortage of takes on AI security. Feeds are full of them — confident, abstract, and untethered from anything that runs. The supply of opinion is effectively infinite, which means opinion by itself is worth almost nothing.
What's scarce is opinion with a working system underneath it. If you've actually built the agent, wired the retrieval, watched the thing fail in a way the blog posts didn't warn you about, and then fixed it — your take is a different kind of object. It carries evidence. It can't be faked without doing the work.
That's the bet this blog makes. Where a piece here makes a claim about how to build or secure these systems, there's usually a real repository behind it that you can open and use however you like. The extra-nerdy five percent will do exactly that, and that's where the interesting conversations start.
The format is the argument: readable for the practitioner who will never open a repo, with a link on the back end for the one who will.