Blog / Writing
Notes from the build.
What I built, what broke, and what I learned along the way.
How I actually sandboxed untrusted code
A few days ago a test pack read my API keys out of a hardened container and shipped them abroad. The container had every flag set and none of them mattered, because the code I was running wanted none of them. So I stopped hardening the room the code runs in and gave it a different room: no network, no keys, no repository. Here is how it is built, and the one case it still doesn't cover.
Read the post · 6 minMy agent hardened the container "perfectly." The API keys walked out anyway.
A coding agent containerized our internal eval tool and did a clean job: read-only filesystem, every Linux capability dropped, no new privileges. It also let any eval pack I run read my API keys and ship them to a server in another country. Both are true at once. The agent hardened the box and never asked who the box was meant to keep out.
Read the post · 7 minMy fine-tuned search model scored better. I'm not using it yet.
I fine-tuned a small on-device search model for voice notes. If search always shows five results, the right note is among them 54.7% of the time, up from 46.0%. Counting only what would appear on screen, it was ahead by less than one point, so I'm not using it yet.
Read the post · 6 minI asked whether ArchByte still worked. Mostly it did, and that was the problem.
ArchByte draws your codebase's architecture with AI agents. I tested it like a stranger would, found what was broken, slow or pretending, and fixed it. Here is what it does today, with real numbers.
Read the post · 8 minFollow along with the Atom feed.