Imam Chowdhury
Short notes on making AI agents, and the systems under them, dependable. Each one is about a single thing worth knowing, in plain words, and many have a working part you can run on this page, on your own device.
A filter that looks for hidden instructions will miss one eventually. What holds is deciding what each agent is able to do, so that a fooled agent has nothing dangerous within reach. Write your own attack below and watch where it stops.
Has a part you can run
An agent that loses power halfway through a payment cannot know whether the payment happened. If it tries again you may pay twice, and if it does not you may not pay at all. One small habit removes the question. Cut the power yourself below.
Has a part you can run
One email when a new note is published: its title, a short summary and the link. Nothing else, and one click stops it.