Glossary
Jailbreak
A prompt crafted to get a model to ignore its safety training or policies and produce output it would normally refuse.
Checks run on our GPUs in Switzerland. Request payloads are not stored.Security
What Jailbreak means
A jailbreak is an input designed to make a model set aside its safety training: role-play framings, obfuscated requests, many-shot examples and similar techniques. Jailbreaks and prompt injection overlap but differ in who is attacking whom. A jailbreak is usually the user trying to get restricted content from the model; an injection is a third party hijacking the model against the user or developer.
Classifiers such as Meta’s Prompt Guard are trained to spot jailbreak and injection attempts in text. That is a different job from judging whether a specific tool call is safe to run. An agent can be jailbroken into a harmful action, but a harmful action can also come from an honest mistake or an invented argument with no jailbreak involved.
MCP Guard judges the action rather than scanning the prompt for attack patterns, so it is complementary to text classifiers rather than a replacement.
Related terms
How Jailbreak connects to the rest of the vocabulary.