CyberGuard dictionary / AI Security
Large Language Model Security
Large language model security is the protection of LLM applications, their data, tools, models, and users against misuse and unintended behavior.
In simple terms
An LLM is often connected to documents, accounts, and actions. Each connection creates a boundary that needs its own access rules and testing.
Technical explanation
A useful assessment covers prompt injection, sensitive information disclosure, data and model poisoning, output handling, supply chain, excessive agency, retrieval controls, and resource limits. Traditional application security still applies.
Risk depends on the whole application: model, retrieved data, prompts, tools, identities, and downstream actions. Security review should test trust boundaries and authorization instead of assuming the model can enforce them.
Examples in practice
A support assistant reads customer files; testing verifies that one customer's prompt cannot retrieve another customer's records.
A customer assistant summarizes account records. Tests verify that retrieved documents are filtered by tenant before the model sees them and that generated text cannot initiate a refund.
Why it matters
The model may produce fluent text while the surrounding application still has serious authorization flaws.
VERIFY & LEARN MORE
Sources and further reading
- genai.owasp.orghttps://genai.owasp.org/llm-top-10/↗
- airc.nist.govhttps://airc.nist.gov/airmf-resources/airmf/↗
Published 2026-10-15 · Last updated 2026-10-15 · CyberGuard editorial content. Source links provide context; they do not imply endorsement.