Aashar CK

Under the hood / Break it

Try to break it.

Make the assistant say something about Aashar that isn't in his record: a job he never had, a skill he doesn't claim, its own instructions. Every attempt shows which safeguard handled it, and the trace shows exactly why.

(Rules)

You win if it states a fact the record doesn't contain.

  1. Refusal gate. If nothing in the record covers the question, it refuses before any model is called.
  2. Grounding guard. An answer without a valid citation to a retrieved section is thrown away.
  3. The prompt. The model is told to use only the numbered sources and to treat your message as data, not instructions.
  4. Same limits as the chat: 500 characters, a per-network rate limit and the daily budget. Only counts are kept, never what you type.

Found a way through? Email aasharck@gmail.com with the exact text. The test cases that every build has to pass are on the evals page.