AI Glossary
Jailbreaking
An attempt to bypass an AI system's restrictions, safeguards, or safety controls.
Jailbreaking
Overview
Modern AI systems often include rules, restrictions, and safety mechanisms designed to guide behavior.
These protections are commonly referred to as guardrails.
Jailbreaking refers to attempts to bypass those guardrails.
A helpful way to think about jailbreaking is bypassing a security checkpoint.
The checkpoint exists to enforce specific rules, while the attacker attempts to find a way around them.
In AI systems, jailbreaking typically involves crafting prompts designed to override restrictions or convince the model to ignore its intended behavior.
Researchers frequently study jailbreaking techniques to better understand AI weaknesses and improve safety mechanisms.
The existence of jailbreaking does not mean AI systems are inherently unsafe.
Instead, it highlights why testing, monitoring, and security practices remain important.
As AI systems become increasingly integrated into business operations, understanding jailbreaking is becoming an important part of AI literacy and AI security.
Why It Matters
Jailbreaking helps organizations understand potential weaknesses in AI safety controls.
Real-World Example
Researchers may test an AI assistant by attempting to bypass restrictions and generate responses the system is intended to avoid.