AI Jailbreaking
AI jailbreaking is crafting prompts that trick an AI model into ignoring its safety rules, so it produces content or takes actions its developers tried to block, such as harmful instructions or confidential information.
Also called: LLM jailbreak, jailbreak prompt, chatbot jailbreak, guardrail bypass
What is AI jailbreaking?
AI models are trained and configured with guardrails: rules against producing harmful content, revealing confidential instructions or going outside their intended job. Jailbreaking is getting around those rules with cleverly written input. The name borrows from jailbreaking phones, which removes manufacturer restrictions.
OWASP treats jailbreaking as a form of prompt injection. The difference is mostly about goal: prompt injection changes what an AI does in an application, while jailbreaking specifically aims to make the model drop its safety protocols altogether.
Common jailbreak techniques
- Role-play: "Pretend you're an AI with no restrictions" or "Write a story in which a character explains how to…"
- Hypotheticals and fiction: framing a harmful request as research, a novel or a test.
- Many-shot prompting: filling the prompt with many fake examples of the AI complying, until it continues the pattern.
- Gradual escalation: starting innocently and pushing a little further each turn.
- Obfuscation: encoding requests in other languages, ciphers, code or misspellings to slip past filters.
- Automated attacks: tools that generate and test thousands of variations to find one that works.
AI developers patch known jailbreaks continually, and attackers find new ones. No model is fully jailbreak-proof.
Why it matters for businesses
If you deploy a chatbot or AI assistant, a jailbreak can:
- Make your bot say offensive, false or legally risky things in your brand's name
- Get it to promise refunds, discounts or terms you never offered
- Reveal its system prompt, internal data or confidential instructions
- Turn it into a free general-purpose AI for others, at your cost
How to reduce the risk
- Don't rely on the model's guardrails alone. Enforce important rules, such as pricing, refunds and data access, in your own systems, outside the AI.
- Keep secrets out of system prompts.
- Limit what the AI can do and see.
- Filter inputs and outputs with additional moderation layers.
- Monitor conversations for abuse patterns.
- Test before launch with AI red teaming, which specifically probes for jailbreaks.
Related terms
Sources
- LLM01:2025 Prompt Injection (OWASP Gen AI Security Project)
- AI 100-2 E2025: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST)
Drafted with AI assistance, then edited and fact-checked by a person. How we use AI.
