Jailbreak (AI) in plain English.

Jailbreaking is the practice of crafting a prompt specifically designed to trick an AI model into ignoring its safety training and producing content it would normally refuse — instructions for harm, explicit material, or anything that violates a provider's usage policy. Common techniques include asking the model to roleplay as an unrestricted alter ego, burying the real request inside a fictional story, or chaining together seemingly innocent steps that add up to something the model would refuse if asked directly. Every major AI lab runs dedicated red-teaming to find and patch jailbreaks before and after release, and the practice is genuinely adversarial: as labs patch known jailbreaks, new ones are discovered within days. In 2026, frontier models like Claude Opus 4.8 and GPT-5.6 are meaningfully harder to jailbreak than their 2023 predecessors, but no provider claims to have solved the problem completely.