Jailbreak
Definition
A jailbreak is an adversarially designed attempt to circumvent security or behavioral boundaries of an AI system.
Attackers can aim with inputs to bypass protection rules, generate unauthorized content, or trigger actions. The terms jailbreak and prompt injection overlap; OWASP classifies jailbreaking as a form of prompt injection aimed at circumventing protection rules.
Simply explained
Like trying to trick a digital security rule through clever wording.
Why it matters in practice
Relevant for publicly accessible AI assistants and agents with tools or enterprise data.
Example from business
A red team tests whether a support bot outputs confidential internal content despite system rules.
Limits and pitfalls
- Only blocking known jailbreak phrases; confusing prompt filters with complete security; granting tool permissions too broadly.
Origin & history
NIST addresses prompt injection and related adversarial attacks in generative AI risk management; security research continuously develops new jailbreak methods.
Use in mid-sized companies
- Possible applications: red-teaming, protection tests, and building multi-layered controls before productive deployment.
Frequently asked questions
How does OWASP define jailbreaking?
As a special form of prompt injection that bypasses protection rules. The terms overlap and are distinguished differently depending on the taxonomy. Can all jailbreaks be blocked? There is no general guarantee.
Sources
- Generative AI Profile · NIST AI 600-1, Generative AI Profile
- OWASP GenAI Security Project · OWASP GenAI Security Project
Sources checked on 10.10.2026.
Editorially reviewed on: