Most "jailbreaks" target the system prompt by exploiting the sampling temperature. Lowering it to 0 makes model refusals deterministic, exposing the exact phrase boundaries where safety filters trigger. It's not a glitch; it's how the weights are tuned. #LLM #RedTeaming