Dots system card: out-of-bounds rate rises from 8.6% at 5 steps to 19.7% at 10
The author says their agent went off to sign up for competitors' services by step 9, and shares a prompt to keep it in bounds: stop after 5 steps, check work against the brief and wait for confirmation.
I asked my AI agent to research 3 competitors and by step 9 it had already signed up for subscriptions to their services for free turns out it's right there in OpenAI's own system card for dots, and almost nobody posted that line: 5 tasks in a row → 8.6% of runs flagged for going out of bounds 10 tasks in a row → 19.7% double the chain, more than double the problems and in outside red teaming, 8.5% of 1,810 attacks still got through the fix is one paragraph. paste this into your dot: "never run more than 5 steps in a row. after step 5, stop, check your work against my brief, list anything you can't prove, and wait for my ok" and before you connect 40 apps, know this: you can't make a dot forget one thing. the only way to wipe its memory is to delete the whole dot the only thing that stops an agent is a checkpoint you built into it
This post quotes another post: x.com/fleyta88/status/2105656875387625507