LLM input is the text or other data provided to a large language model, including prompts, instructions, and context assembled by an application or AI agent. It may be processed, logged, or retained according to the system’s design, so it should not automatically be treated as private storage.
In cybersecurity, inputs matter because they can carry passwords, API keys, access tokens, or other sensitive information into systems with different permissions and retention practices. An agent might retrieve a credential and pass it into a model prompt while completing a task, creating exposure even without an attacker. Reduce risk by minimizing input data, redacting secrets, limiting agent access, and avoiding sending credentials to a model unless strictly necessary.