AI & LLM Security
AI/LLM Security & the OWASP LLM Top 10
Why AI apps are a whole new attack surface, and the risks that define it.
Every company is bolting LLMs onto their products — chatbots, copilots, agents, RAG search. These systems take untrusted natural-language input and often have access to tools, data, and actions. That makes them a rich, immature attack surface, and a booming specialization.
The OWASP Top 10 for LLM Applications
- LLM01 Prompt Injection — untrusted input overrides the model's instructions (the core bug).
- LLM02 Sensitive Information Disclosure — leaking system prompts, secrets, or training data.
- LLM05 Improper Output Handling — trusting model output that then hits XSS/SQLi/command sinks.
- LLM06 Excessive Agency — an agent with tools/permissions doing harmful actions.
- LLM04 Data & Model Poisoning, LLM08 Vector/Embedding Weaknesses, and supply-chain risks.