Glossary
Jailbreak
A jailbreak is an attempt to make the model ignore guardrails and the system prompt. It is related to prompt injection, but often arrives as adversarial conversation.
Also called دورزدن محدودیت
Defense belongs in the deterministic layer, not in a polite prompt. If a friendly tone can break policy, there is no policy.
Let’s make your processes agentic
Where do we start?