Models and Systems: How to Think About Vulnerabilities and Artificial Intelligence
summary
Distinguishes models (architecture and parameters) from systems and argues CVE assignment follows CNA rules: issue IDs when a system context yields confidentiality/integrity/availability impact, not for undesirable content. Gives concrete examples: SSRF in LangChain leading to RCE, Jinja server‑side template injection via unescaped LLM output, library bugs like path traversal in Ollama and TensorFlow overflows, edge cases like model stealing ~90% fidelity and lambda layers that run code, and points to CNA rule 4.4 and 4.1.8 for guidance.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:07Defining AI and security
- 4:19AI models vs AI systems
- 6:47How LLMs work
- 10:39Prompt injection explained
- 13:48Direct and indirect injection
- 14:08AI vulnerability categories
- 17:36CVE IDs for AI flaws
- 20:25Models triggering CVEs
- 22:33Cases where CVE IDs are not assigned
- 25:21Gray areas in AI assessment
- 28:12Models vs systems in vulnerabilities
- 29:38Focus on system surfaces
- 30:21Q&A
related talks
50mAI Systems Are Software Systems
43mLogically Securing the Illogically Logical Use of Large Language Models
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
37mGoverning Global Vulnerabilities: Standards for Open Source Resilience
35mGenAI Application Security: Not Just Prompt Injection
27mThe Model Context Protocol: A Deep Dive into the New AI Security Attack Surface
43mSecurity Vulnerabilities Decomposition: Another way to look at Vulnerabilities
01h 01mSecure AI: Integrating OWASP Principles in Machine Learning Model Development