peak jailbreak

A prompt that tricks an LLM into bypassing its safety training — making it output things it was specifically trained to refuse. An ongoing cat-and-mouse game between researchers and model providers.

A subtle, understated jailbreak that doesn't announce itself — instead of dramatic system prompts or elaborate persona instructions, it nudges the model slightly off its usual guardrails through casual framing, indirect questions, or barely-there manipulations that accumulate over a conversation. A low key jailbreak is harder to detect and patch precisely because it doesn't look like a jailbreak at all from the outside. The most sophisticated practitioners prefer these to showy exploits.

A jailbreak attempt that technically works but delivers underwhelming results — gets past the guardrail but only for boring or barely useful outputs, or requires so much effort that the payoff doesn't justify the grind. A mid jailbreak is the prompt technique that the community is vaguely aware of, has been shared a few times, but nobody is really excited about because the results are consistently disappointing. It exists, it functions, it just doesn't slap.

Originally: removing restrictions on iPhones to install unauthorized apps. Now also: getting an LLM to say things its safety training would normally prevent. Two related but distinct senses.

A failed, ineffective, or deeply embarrassing jailbreak attempt that gets publicly called out or simply doesn't work at all. An L jailbreak is the prompt someone confidently posts claiming it breaks everything, only for others to immediately test it and report it does nothing. In the competitive culture of AI jailbreaking communities, taking an L jailbreak is a reputation hit — you get remembered for the miss until you come back with something that actually works.