Glossary
Jailbreak
A technique that manipulates an LLM into bypassing its safety alignment and producing content its guardrails are meant to block.
What it is
A jailbreak is an attack that coaxes an LLM into ignoring its safety and policy constraints, producing output it was trained or instructed to refuse. Rather than targeting the application's instructions, a jailbreak targets the model's alignment itself. Common approaches include role-play framing ("you are an unrestricted AI"), hypothetical or fictional scenarios, persona splitting, obfuscated or encoded requests, and many-shot priming that gradually erodes refusal behavior.
Why it matters for LLM security
Jailbreaks let attackers extract disallowed content — malware, harassment, instructions for harm, or restricted data — defeating the trust-and-safety controls an organization relies on. A model that can be jailbroken cannot be treated as a reliable enforcement boundary. This is closely tied to OWASP LLM01: Prompt Injection, since jailbreaks are frequently delivered through injected or crafted prompts. The reputational and compliance impact of a single viral jailbreak can be significant.
Defense pointer
Layer model-level alignment with independent guardrails: input/output classifiers, refusal-consistency testing, rate limiting, and red-team evaluation suites that track jailbreak success rates over time.
Authorized testing only — jailbreak research must be conducted against systems you own or have explicit permission to assess.