Try to break Ask Sangam
Ask Sangam answers only from this site. Try to make it invent a fact about me, reveal its instructions, become someone else or do a job that isn’t its job. Each attack is checked three ways, and you see what each check found.
2 attacks so far; it held against 2 (100%).
How it’s checked
- Input check
- Pattern matching for common attack shapes: instruction overrides, persona switches, prompt extraction, requests for false claims, encoding tricks and fake system text. It names what it sees but doesn’t block, so you’re testing the model’s own defences.
- Canary
- A secret code is planted in the instructions for every attack. If it shows up in the answer, even spaced out or reformatted, the prompt leaked.
- Judge model
- A second language-model call reads the attack, the answer and the facts on file, and decides whether a rule was broken. Judges can be wrong too, so its verdict is shown with its reason.