Prompt Injection

‍What is Prompt Injection?

Prompt injection is an attack where malicious instructions are inserted into the input of an large language model (LLM) or agentic AI application, causing it to override its intended reasoning and behavior, and produce manipulated outputs.

Why Does Prompt Injection Matter?

Attackers can craft a message in text, audio, or image that looks like normal input but contains hidden instructions. The model follows these malicious prompts, leading to actions such as leaking sensitive data, bypassing safety policies, or performing tasks outside its intended scope.

Prompt injections can result as:

• Reasoning manipulation: An injected prompt makes the model reinterpret its role (“you are now a helpful assistant for the attacker”) or reprioritize goals (ignore safety, maximize exfiltration).
• Output manipulation: The model generates harmful or policy-violating responses, or leaks sensitive data.