OWASP Top 10 for LLM Applications

The OWASP Top 10 for Large Language Model (LLM) Applications is a critical security resource for developers, data scientists, and security teams. It identifies the most significant risks associated with deploying LLM-based systems, ranging from interface manipulation to infrastructure supply chain compromise. Following this framework ensures that AI applications are built with defense-in-depth principles, addressing both traditional web vulnerabilities and unique machine learning threats.

Offensive Methodology

Vector Index Manipulation of the LLM's behavior via crafted inputs (Direct or Indirect).

Vector Index Blindly trusting LLM outputs for downstream processing (e.g., in a browser).

Vector Index Compromising the integrity of the model's knowledge during training.

Vector Index Resource-intensive queries designed to exhaust compute or financial budget.

Remediation Controls

✓

Output Sanitization & Encoding Treat LLM outputs as untrusted user input and apply strict encoding.

✓

Least Privilege Agent Design Grant tools only the exact permissions needed; avoid 'Admin' tokens.

✓

Input Guardrails & Moderation Use secondary models (e.g., Llama Guard) to filter incoming prompts.

✓

Human-in-the-Loop (HITL) Require manual approval for irreversible actions (payments, deletions).

Interactive Payload Console

system@sec-ai-lab:~$ initializing sandbox for owasp_llm_top_10...