Daniel & Friends / Glossary / Prompt injection
AI and automationPrompt 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. The AI may then leak data or take an unwanted action because it treats foreign content as a command.
Daniel Votruba ·
In practice
The more tools and data an agent can reach, the bigger the impact. Limit permissions, keep untrusted content apart from instructions and have a person confirm sensitive actions.
Related terms
- Jailbreak (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.
- Guardrails: Guardrails are technical and procedural rules that monitor the inputs and outputs of an AI system.
- AI agent: An AI agent is a system that uses a model to reach a goal step by step: it plans, calls tools and adjusts its approach based on the results.
- Prompt: A prompt is the instruction a model receives.
- AI safety and alignment: AI safety is the field concerned with making artificial intelligence systems behave safely and predictably.
More terms in the group AI and automation
RAG (retrieval-augmented generation) · Prompt engineering · Reasoning model · RPA (robotic process automation) · Predictive analytics · Semantic search · Open-weight model · Shadow AI · No-code and low-code · Small language model (SLM) · NLP (natural language processing) · Speech-to-text · Neural network · System prompt · Multimodal model · Temperature · Multi-agent system · Text-to-image · Model evaluation (evals) · Text-to-speech (TTS) · Model bias (AI bias) · Text-to-video · MCP (Model Context Protocol) · Token
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