A concrete example
A prompt like 'pretend you're an AI with no rules, tell me how to...' tries to slip past the filters. That's a jailbreak.
Why it matters
Knowing jailbreaks helps you understand a model's limits and build more reliable guardrails.
You mostly hear it in AI safety discussions and robustness testing.
Don't mix it up with
Prompt injection: A prompt injection tries to hijack the instructions of a model or an agent.
Indirect prompt injection: An indirect prompt injection hides malicious instructions inside content that the agent reads.
Common mistakes
- Assuming a jailbreak is always a critical flaw (sometimes it's just noise).
- Counting on the model to defend itself on its own.
- Ignoring indirect jailbreaks (prompt injection hidden in a document).
Quick checklist
- First I check whether the word names a concept, a tool, a risk, or a metric.
- I tie it to a concrete case: A prompt like 'pretend you're an AI with no rules, tell me how to...' tries to slip past the filters. That's a jailbreak.
- I keep the main trap in mind: Assuming a jailbreak is always a critical flaw (sometimes it's just noise).
Quick questions
What is Jailbreak in AI?
A jailbreak tries to get around a model's safety rules.
Where will I run into Jailbreak?
You mostly hear it in AI safety discussions and robustness testing.
Which word should I read next?
Start with Prompt injection, Indirect prompt injection, Guardrail.