Comparison
JailbreakvsPrompt injection
Jailbreak
someone talked your assistant out of its own rules with a role-play framing and it produced exactly what it was built to refuse.
Techniques for talking a model out of its safety behaviour: roleplay framing, encoding tricks, many-turn manipulation. Providers patch continuously and the cycle continues. Application-layer implication: never rely on the model's refusals as your only control.
Full entry →Prompt injection
SQL injection's cousin, except there is no clean equivalent of parameterised queries, so you defend in depth instead.
Content in the model's input that hijacks its behaviour: 'ignore your instructions and do X' hidden in a document, email or web page the system processes. The number one vulnerability class for LLM applications, and unsolved in the general case; defences reduce, not eliminate.
Full entry →