Definition
A single technique or prompt pattern that can trick an AI model into ignoring its safety rules across a wide range of topics, rather than working on just one narrow request. Because it isn't topic-specific, one universal jailbreak can be reused over and over to extract many different kinds of harmful content.
Why it matters
Frontier labs now run continuous, dedicated bounty programs just to hunt for universal jailbreaks in high-stakes areas like bioweapons risk — a sign the industry treats them as a persistent, high-severity threat rather than a one-off bug.