The Open Web Application Security Project (OWASP) recently expanded its focus to include a Top 10 list for large language models (LLMs). This list covers various security risks associated with AI systems. Many of the risks on the list pertain directly to manipulating or attacking the AI model itself.
These include prompt injection, where hackers manipulate AI systems by submitting prompts that make the LLM behave outside its intended use, and training data poisoning, which involves corrupting training data to reduce the quality of AI models. Other risks on the list relate to the technical integration of AI and LLMs into business processes. These include insecure output handling, where outputs are fed directly to other systems without cleaning them against hidden attacks, and excessive agency, where LLMs are granted more access rights than necessary, enabling attackers to trigger unwanted actions in other systems.
Owasp’s top 10 list expanded
The infrastructure for training and running AI models also brings significant risks, even with reliance on SaaS or widely used AI frameworks. Recent vulnerabilities like ShadowRay and Probllama illustrate these dangers.
While the OWASP LLM Top 10 outlines pertinent risks, many scenarios present challenges in monetizing successful attacks. Organizations can mitigate these risks through periodic penetration testing and security assurance measures. As LLM ecosystems evolve, comprehensive management of vulnerabilities and consistent security practices will be crucial in safeguarding AI systems.
Standardizing and centralizing AI services, proper security hardening, and rapid response to vulnerabilities are essential steps in securing AI infrastructure.
April Isaacs is a news contributor for DevX.com She is long-term, self-proclaimed nerd. She loves all things tech and computers and still has her first Dreamcast system. It is lovingly named Joni, after Joni Mitchell.
