Infosec glossary

Prompt Injection

Prompt injection became a core AI security term because language-based systems can be manipulated in ways that look different from traditional input handling problems. The term is useful when it clearly describes instruction override, context manipulation, and unintended model behavior.

Threat language 6 min read By Infosec Writing Studio editorial team Published Apr 18, 2026

What is Prompt Injection?

Prompt injection is an attack in which an adversary provides crafted input that causes an AI system or language model to ignore, override, or manipulate its intended instructions, often leading to unsafe behavior, data exposure, or unintended actions.

Key takeaways

Prompt injection is an AI-specific attack pattern, not just a generic “bad prompt.”
The core issue is instruction manipulation inside language-based systems.
The term often includes both direct and indirect prompt injection.
If the explanation treats it as simple user misuse, it misses the attack logic.

Common mix-ups

Treating prompt injection as the same thing as jailbreaking in every case.
Explaining prompt injection without mentioning instruction override or context manipulation.
Assuming the risk only exists in chatbots and not in AI-enabled workflows or agents.

Why prompt injection matters

Language models process instructions and data through the same general medium: text or semantically interpretable input. That creates a unique problem where hostile or misleading input can compete with or override intended instructions.

This is why prompt injection matters in AI security. The issue is not only model output quality. It is the security risk created when the system treats malicious input as valid instruction context.

How the attack is usually described

Prompt injection is usually described through direct or indirect instruction manipulation. In a direct case, the adversary inserts instructions openly. In an indirect case, the malicious input is embedded in content the system later processes, such as documents, web pages, or workflow inputs.

That distinction helps the term stay precise, especially in agent and retrieval-connected systems where the model is consuming outside content as part of the task.

Direct instruction override
Indirect prompt manipulation through content
Unintended data or behavior exposure
Risk in agents, assistants, and retrieval workflows

How to explain prompt injection clearly

The cleanest explanation of prompt injection names the instruction problem first. The attack is about getting the system to follow the wrong semantic path, not just about “asking a bad question.”

That also helps separate prompt injection from broader AI risk categories such as model misuse, governance gaps, or unsafe output quality.

The term is clearer when the nearby language is clear too.

Use the pages below when you need adjacent terms, category context, or a longer explanation instead of leaving the definition to stand on its own.

Adjacent terms

AI Security Secure by Design Software Supply Chain Security

Sources used to check the definition and terminology

OWASP reference for Prompt Injection

Open the external source used to verify the term's scope and related language.

Guides

Where the definition expands into a longer explanation

How to Review Cybersecurity Content Before Publish

A practical pre-publish review process for cybersecurity content covering terminology, claims, audience fit, proof, structure, and trust.

Cybersecurity Content Examples: What Good Looks Like

A practical guide to the traits that separate strong cybersecurity content from generic B2B copy, with examples of what to aim for across common asset types.