Security risks to plan for when deploying LLM agents

LLM services can leak secrets from prompts, context, or training data. Prompt injection and jailbreaks can push the model to ignore its intended rules. Agents with tools or plugins create a bigger risk because they can move from answering questions to changing real systems.

RAG pipelines can also bring in third-party content that hides harmful instructions. Long sessions may cause agents to collect too much sensitive context, and model behavior can change over months even when speed and uptime still look normal. In LLM apps, the line between untrusted input and internal logic can blur quickly.

A safer design treats everything before the model as potentially hostile, labels where information came from, and limits what models and agents can see or do.