Steve Wilson and Gavin Klondike -- OWASP Top Ten for LLM Release

The first OWASP Top 10 for Large Language Model Applications gave developers and security teams a shared threat model for a rapidly changing technology. Project leaders Steve Wilson and Gavin Klondike walk through the inaugural list and explain why familiar controls need to be reconsidered around probabilistic systems. They cover prompt injection, insecure output handling, training-data poisoning, denial of service, supply-chain vulnerabilities, sensitive information disclosure, insecure plugin design, excessive agency, overreliance, and model theft. Along the way, they distinguish application risks from attacks on the underlying model and show how plugins and autonomous actions can enlarge the blast radius. The episode closes with what the project learned from community feedback and how the list would evolve as real-world LLM deployments exposed new failure modes.

The Application Security Podcast is brought to you by Security Journey.

About Security Journey
Security Journey provides application security education for developers and everyone in the software development lifecycle.
→ Learn more about Security Journey

Connect with Steve Wilson and Gavin Klondike:
→ Steve Wilson on LinkedIn
→ Gavin Klondike on LinkedIn
→ OWASP GenAI Security Project

Mentioned in this episode:
→ OWASP Top Ten for LLM Applications project homepage
→ OWASP Top 10 For LLMs 2023 Slides V1 1
→ ChatGPT
→ LangChain
→ Tay (Microsoft chatbot)
→ DEF CON

Follow the Application Security Podcast:
➜ Home
➜ X
➜ LinkedIn
➜ YouTube
➜ Instagram
➜ Facebook

Chapters:
00:00 Introducing the OWASP Top 10 for LLM Applications
01:37 A threat model for large language models
05:39 Risk 1 — Prompt injection
09:00 Risk 2 — Insecure output handling
11:35 Risk 3 — Training-data poisoning
15:14 Risk 4 — Model denial of service
19:11 Risk 5 — Supply-chain vulnerabilities
28:16 Risk 6 — Sensitive information disclosure
33:28 Risk 7 — Insecure plugin design
38:27 Risk 8 — Excessive agency
42:39 Risk 9 — Overreliance
46:31 Risk 10 — Model theft
49:56 Lessons for the next release