Daniel & Friends / Glossary / Jailbreak (AI)
AI and automationJailbreak (AI)
A jailbreak is an attempt to make an AI model bypass its own safety restrictions, typically through a cleverly worded request, role play or a gradual push past the limits. The goal is an answer the model would normally refuse.
Daniel Votruba ·
In practice
If you deploy your own chatbot, test it for such attempts before launch. Enforce forbidden topics outside the model as well, for example with an output filter.
Related terms
- Prompt injection: Prompt injection is an attack in which a malicious instruction hidden in a text, an email or a web page overrides the model’s original task.
- Guardrails: Guardrails are technical and procedural rules that monitor the inputs and outputs of an AI system.
- AI safety and alignment: AI safety is the field concerned with making artificial intelligence systems behave safely and predictably.
- Large language model (LLM): A large language model is a model trained on a huge amount of text that predicts the next words and can write, summarise and answer.
More terms in the group AI and automation
Inference and latency · Machine learning (ML) · Human in the loop · MCP (Model Context Protocol) · Hallucination · Model bias (AI bias) · Model evaluation (evals) · Grounding · Multi-agent system · Generative AI · Multimodal model · Function calling (tool use) · Neural network · Foundation model · NLP (natural language processing) · Fine-tuning · No-code and low-code · Few-shot and zero-shot prompting · Open-weight model · Explainable AI (XAI) · Predictive analytics · EU AI Act · Prompt · Embedding
Terms in practice
Need help putting them to work?
Terms are one thing, putting them to work is another. We go through which of them pay off for you and say so plainly. You get a first opinion for free and we reply within 24 hours.
Book a call