Prompt injection

Crafting input so that a language model follows an attacker's instructions instead of its intended ones.

Covers direct and indirect prompt injection, where an attacker's instructions are typed in or hidden in content the model reads, such as web pages or documents. Corrupting the data a model learns from is Data poisoning; injected input that runs as program code or operating system commands is Code injection or Command injection.

Scheme

• Technologies and capabilities

Broader terms

• Adversarial machine learning
• Injection flaws