OWASP LLM 应用程序前 10 名：新增内容、变化以及如何保持安全

Taking a proactive approach to securing your software requires that you stay at the forefront of the latest standards and compliance requirements. After all, the cybersecurity landscape is constantly in motion with new threats and vulnerabilities, especially as new technologies emerge. Never has this been more true than today where we collectively sit at an inflection point of AI where new evolutions and use cases seem to emerge every day. 

To address these challenges, The OWASP Foundation recently released its updated version of the OWASP Top 10 for Large Language Model (LLM) Applications, which seeks to inform developers, architects and other contributors to software delivery of the potential risks when deploying LLMs and Generative AI applications. And we at Secure Code Warrior are excited to announce that the changes and updates in this latest version are already implemented and available in our secure code learning platform. With these newly available and updated materials, all of our users can stay at the forefront of risk mitigation when utilizing LLMs.

What is new in this update?

OWASP has removed two items from its previous Top 10:

• Insecure Plugin Design - which relates to how LLMs interact with plugins and how plugins interact with external storage or services. 
• Model Theft - referring to unauthorized replication or acquisition of machine learning models or AI system.

In accordance with previous versions of the OWASP Top 10, Secure Code Warrior had Guidelines associated with these vulnerabilities as part of our LLM Top 10 Course. These Guidelines, which provide digestible information regarding vulnerabilities and security concepts in an easy-to-understand, readable format, have since been removed from the Course curriculum. However, the guidelines remain available in Explore, along with all the other learning materials we have to offer. 

Keeping its Top 10 to an official 10, OWASP has added two new items:

• System Prompt Leakage - where typically hidden prompts that guide a model's behavior get exposed to users.
• Vector and Embedding - which can expose specific, proprietary, or real-time information that is not publicly available

Guidelines for these vulnerabilities have been added to the LLM Top 10 Course, and just like the Guidelines that were removed, these two are also accessible in Explore for users that want to take advantage of self-paced learning.

Lastly, OWASP also made some changes to existing vulnerability categories in its list, renaming some categories to be more expansive or specific, and making modifications to their definitions. Our Guidelines supporting these topics have now been updated to reflect both the minor changes from OWASP’s guidance as well as the new naming conventions. Additionally their listing in order of priority has been updated to match the order set in the OWASP LLM Top 10.

At Secure Code Warrior, we’re committed to helping our users stay ahead of the curve. With the latest OWASP updates already reflected in our agile learning platform, we’ve made it easy for our users to access up-to-date training materials that cover the most current vulnerabilities and mitigate risk when deploying LLM and Generative AI technologies. Whether you’re navigating the newly introduced threats of System Prompt Leakage or Vector and Embedding, or updating your understanding of Misinformation and Unbounded Consumption, our platform provides the resources you need to master these critical concepts improving your security posture.