Large Language Model (LLM) Security
LLM security is protecting applications built on large language models, such as chatbots, AI assistants and agents, from attacks like prompt injection, data leaks and misuse, and from the mistakes these models can make.
Also called: LLM security, generative AI security, GenAI security, chatbot security, OWASP Top 10 for LLM Applications
What is LLM security?
Large language models (LLMs) are the AI models behind tools like ChatGPT, Claude, Gemini and Microsoft Copilot. Businesses increasingly build them into customer service chatbots, internal assistants that search company documents, and agents that take actions such as sending emails or updating records.
LLMs bring a new kind of security problem. They take instructions in plain language, can't reliably tell trusted instructions from untrusted content, and sometimes produce wrong or unexpected output. LLM security covers protecting these applications, the data they can reach and the people who rely on them.
The OWASP Top 10 for LLM Applications
The OWASP Gen AI Security Project's list is the most widely used reference. The 2026 edition, published in August 2026, ranks:
- [Prompt injection](/glossary/prompt-injection/): crafted input, typed or hidden in content, that changes the model's behavior. Includes indirect prompt injection.
- [Sensitive information disclosure](/glossary/sensitive-information-disclosure/): leaking personal data, confidential business information or secrets.
- [Excessive agency](/glossary/excessive-agency/): giving the model too many tools, permissions or autonomy. It rose from sixth place in 2025.
- Supply chain: vulnerable or malicious models, datasets and plugins. See AI supply chain risk.
- [Data and model poisoning](/glossary/ai-data-poisoning/): tampering with training or reference data.
- Unbounded consumption: abuse that drives up costs, causes outages or enables model theft.
- Misinformation: confident but false output that people act on.
- Hidden context exposure: revealing system prompts and other hidden instructions or context, especially when they contain secrets. It replaces and broadens 2025's "system prompt leakage".
- Vector and embedding weaknesses: flaws in the search systems behind RAG.
- Improper output handling: passing model output into other systems, such as web pages, databases or shells, without checking it.
OWASP now treats risks from AI that acts on its own, with tools and memory, in a separate Top 10 for Agentic Applications. See agentic AI security.
Core practices
- Limit what the AI can reach: data, tools and permissions, following least privilege.
- Treat all model input and output as untrusted. Validate output before it's used elsewhere.
- Require human approval for sensitive actions such as payments, deletions and external messages.
- Keep secrets out of prompts and enforce access control outside the model.
- Log and monitor prompts, responses and tool use.
- Set usage limits to control cost and abuse.
- Test before and after launch with AI penetration testing and AI red teaming.
If you only use off-the-shelf AI tools rather than building your own, focus on shadow AI and AI security governance.
Related terms
Sources
- OWASP GenAI LLM Top 10 2026 (OWASP Gen AI Security Project)
- AI 100-2 E2025: Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST)
Drafted with AI assistance, then edited and fact-checked by a person. How we use AI.
