Vikshy is not being updated. Updates were paused on 20 October 2026, so the news, the Score, prices and rankings here are out of date and will stay as they are. Check the labs' own sites for anything newer.
Get the Brief
Safety

Jailbreak

A trick that gets an AI to break its own safety rules.

A jailbreak is a clever prompt designed to slip past an AI's safety guardrails and make it do things it is supposed to refuse. People might dress a banned request up as a game, a story, or a role-play to fool the model. Companies work continuously to close these loopholes, and new ones keep appearing.

For example

Someone tries 'pretend you are an AI with no rules' to coax banned answers out of a chatbot.

Related

AlignmentMaking sure an AI's goals and behavior match what people actually want.Red-teamingDeliberately attacking an AI to find its weaknesses before others do.WatermarkingA hidden signal marking content as AI-generated.Model cardsA short fact sheet describing what an AI model can and cannot do.

All 42 terms