OWASP has released the 2026 edition of its Top 10 for Large Language Model (LLM) Security, with prompt injection retaining the number one spot for the second consecutive year. The updated framework adds three new categories: LLM DoS (resource exhaustion attacks), supply chain vulnerabilities in AI models, and excessive agency (models taking unintended actions).
The framework has gained widespread adoption across the industry, with major cloud providers and AI platforms incorporating OWASP LLM categories into their security review processes. The report notes that while awareness has improved significantly, many organizations still lack basic guardrails for their LLM deployments.
Key recommendations include implementing input validation and sanitization for LLM prompts, restricting model permissions to the minimum required, conducting regular red-teaming exercises, and maintaining comprehensive audit logs of model interactions.