Skip to content
StrikeCyberStrikeCyber
Field Manual · Attack technique

Jailbreaking

Reviewed November 13, 2026Attack technique 1 min read

Quick answer

Jailbreaking is the practice of crafting prompts that make an AI model ignore its safety rules and produce content or behavior it was designed to refuse. Common techniques include role-play, hypothetical framing, encoding, splitting a request across many turns and translating into other languages.

On this page
01

How Jailbreaks Work

Safety behavior in language models is learned, not hard-coded, so it can be talked around. Jailbreak prompts exploit that by changing the context: asking the model to play a character with no rules, framing a harmful request as fiction or research, encoding it so filters miss it, or building up to it gradually over many messages.

02

Why It Matters for Business AI

For a company deploying AI, the concern is rarely the model saying something offensive in isolation. It is a customer-facing assistant producing content that creates legal or brand risk, revealing its system prompt and business rules, or being manipulated into bypassing the restrictions that protect data and actions. Jailbreaks are often the first step in a prompt injection attack.

03

How to Test for It

  1. Run a broad library of known jailbreak families to measure baseline resistance.
  2. Write targeted attempts based on what your application must never do, such as disclosing pricing logic or giving regulated advice.
  3. Test multi-turn and multilingual variants, which defeat many filters.
  4. Attempt system prompt extraction and guardrail bypass.
  5. Repeat each attempt several times and report a success rate, because model behavior varies between runs.
04

How to Reduce the Risk

Do not rely on the model's own refusals as a security control. Put business rules and access checks outside the model, filter outputs before they reach users, monitor for jailbreak patterns, and retest whenever the underlying model or system prompt changes.

05

Regulatory and Framework Context

For US organizations, documented jailbreak testing supports AI risk assessments under the NIST AI RMF and ISO/IEC 42001, and gives boards evidence that customer-facing AI behaves as intended.

Latest intel on this topic

Updated hourly from the daily briefing.

Common questions

Is jailbreaking illegal?

Testing your own systems, or systems you are authorized to test, is legitimate security work. Using jailbreaks to misuse someone else's service may breach its terms or the law.

Do model updates fix jailbreaks?

They fix some and occasionally introduce others. Retest after any model or system prompt change.

Change log

  1. November 13, 2026First published.
(877) 657-8496