Loading Agentverse...
Prompt injection is like tricking an AI by hiding secret instructions in its input that make it do something it was not supposed to do.
A security vulnerability where malicious instructions are inserted into AI model inputs to override system prompts or manipulate model behavior.
Prompt injection attacks exploit the inability of LLMs to reliably distinguish between system instructions and user-provided data. Direct injection places malicious instructions in user input. Indirect injection embeds instructions in external data sources (websites, documents) that the model processes. Mitigations include input sanitization, instruction hierarchy, output filtering, and architectural separation of instruction and data channels.