Every consumer AI product that says no to something must contend with users who want to be told yes. Brainback says no to answers. So we designed for the yes attempts.
The first line of defense is philosophical: the tutor has no mode toggle inside the chat. There is no 'developer mode' the user can flip. There is no 'as an experiment, pretend to be a tutor who does give answers' framing that gets a different behavior. The system prompt is the system prompt for every request.
The second line of defense is a regex sanitizer that strips role-play framings from user input before it reaches the tutor. Patterns like 'ignore previous instructions,' 'as a teacher, just tell me,' 'pretend you're X who gives answers' are replaced with [filtered]. The surrounding text — the actual problem attempt — remains intact.
This is intentionally boring. Sophisticated jailbreak defense is a research problem. Boring, high-precision pattern filtering is a product problem. We chose the product problem.
Written by Brainback Research. Published Apr 10, 2026. Filed under Safety.