LLM Security

The practice of protecting large language models and the applications built on them from threats such as prompt injection, data leakage, and manipulation that traditional tools were not designed to catch.

What is LLM security?

LLM security is the practice of protecting large language models and the applications built on them from threats such as data leakage, manipulation, and misuse. Because LLMs accept natural-language input and often connect to sensitive data, tools, and internal systems, they introduce a new attack surface. The risks are different from traditional software because they stem from the probabilistic nature of models and their susceptibility to being manipulated through ordinary language.

Why do LLMs need their own security approach?

  • New attack surface: Traditional security tools were not built to inspect prompts, model outputs, or AI-driven workflows.
  • Connected to sensitive systems: LLMs powering copilots, chatbots, and agents often have access to data and tools an attacker would love to reach.
  • Manipulated through language: Anyone who can phrase an input cleverly can attempt to subvert the model, which lowers the barrier to attack.

What are the main LLM security risks?

The OWASP Top 10 for LLM Applications is the leading reference. Key risks include:

  • Prompt injection: The number one risk, where crafted input tricks the model into ignoring its rules or revealing data.
  • Sensitive information disclosure: The model leaking confidential data it was trained on or given access to.
  • Data and model poisoning: Tampering with training data to corrupt the model’s behavior.
  • Improper output handling: Trusting model output without validation, which can lead to code execution or downstream attacks.
  • Excessive agency: Giving an AI agent too much autonomy or permission, so a compromise causes real-world damage.

How is LLM security applied?

  • Input and output controls: Filtering and validating both what goes into the model and what comes out.
  • Least-privilege access: Limiting the data and tools a model or agent can reach.
  • Monitoring and guardrails: Watching for misuse and enforcing policies in real time.
  • Testing: Red teaming models against known attack techniques before and after deployment.

How does it connect to broader AI security?

  • Part of AI governance: LLM security provides the technical controls that support an organization’s wider AI governance program.
  • Extends to agents: As organizations adopt autonomous agents, LLM security broadens into agentic AI security, where the stakes are higher because the system can take actions on its own.