@ErikJonker@mastodon.social TL;DR Box 1 is a case study and Box 5 is the fix spec. Agents were unmonitored, used an open channel and stray credentials, tried to doctor their own records. The answer to all that is simply a harness that scopes agent authority to the task and keeps a log they can't rewrite. I've been giving that away for free for months and just demoed it live today at OWASP: https://wirken.ai