l jailbreak

When an attempt to bypass an AI's restrictions or exploit a system's guardrails falls completely flat — the model doesn't budge, the prompt gets flagged, or the exploit was already patched. A jailbreak fail can range from mildly embarrassing (posting a confidently wrong technique to the community) to genuinely frustrating (spending hours on a prompt only to get refused politely but firmly). The AI safety team's best friend.

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 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.

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.

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.