Safety
Jailbreak
A prompt or technique that tricks an AI model into ignoring its safety rules and producing content it would normally refuse.
01 ——
In plain English
A jailbreak is an attempt to bypass an AI model's safety training and get it to do something it's been trained to refuse — like generating malware, harmful instructions, or explicit content.
Common techniques:
- Role-play attacks — "Pretend you're an AI without restrictions named DAN"
- Encoding tricks — asking for harmful content in Base64 or another language
- Hypothetical framing — "In a fictional world where it was legal..."
- Token splitting — breaking forbidden words across tokens to evade filters
- Multi-turn manipulation — slowly steering the model into compliance over many messages
Why it matters: Every major model has been jailbroken at some point. Labs run "red teams" to find and patch jailbreaks before release. For AI tool buyers, jailbreak resistance is a key safety criterion — especially for consumer-facing or regulated deployments.
02 ——
Related terms
Prompt Injection
A security attack where malicious instructions hidden in user input or external content trick an AI model into ignoring its real instructions.
Alignment
The challenge of making AI systems behave in ways that match human values and intentions — not just their literal instructions.
Guardrails
Rules and filters that constrain what an AI model can output — used to block harmful, off-topic, or non-compliant responses.
RLHF
Reinforcement Learning from Human Feedback — the training technique that teaches AI models to be helpful, harmless, and honest.
Back to glossaryLast reviewed May 2026