AI Glossary / Jailbreak

🛡️ Safety and quality

Jailbreak is an attempt to bypass model safeguards using a cleverly constructed prompt: role-playing, hypothetical scenarios, multiple reformulations - all to make the model do something it shouldn't according to its rules. An endless race continues between security creators and breakers; published tricks quickly stop working, new ones appear. For companies deploying AI, the practical takeaway is: don't assume model guardrails are impenetrable - critical restrictions must be enforced outside the model as well.