Logically Securing the Illogically Logical Use of Large Language Models
summary
Scenario: someone downloads a model from Hugging Face and runs it on internal data; scanning reveals instances where models carry a malware backdoor found and open shells. The remedy is concrete: apply NIST 800-53 rev5 and AI RMF 1.0 controls—access control (least privilege for model access), configuration management (scan and harden), incident response (triage AI-specific breaches), supply chain checks (S2C2F, verify Ray/other components for CVEs) before deployment.
AI-written summary · about
chapters (13)
- 0:00Securing LLMs
- 2:29Speakers and agenda
- 6:20Security fundamentals for LLMs
- 7:22LLM innovation risks
- 11:17Risk ownership in LLMs
- 15:15LLM security challenges
- 19:33NIST 800-53 for LLMs
- 27:43AI Risk Management Framework
- 35:35Protecting LLMs with frameworks
- 36:18Scanning downloaded models
- 38:01Zero trust for AI workloads
- 40:53Preemptive LLM risk management
- 42:40Q&A
related talks
28mFlawed ML Security: Mitigating Security Vulnerabilities in Data & Machine...
28mSecurity of Machine Learning
45mSecuring Software Supply Chain: Problems, Solutions, and AI/ML Challenges
31mSecurity Considerations for Services Using AI Models
47mA Step-by-Step Guide to Securing Large Language Models (LLMs)
29mThreat Modeling Meets Model Training: Web App Security Skills...
50mBuilding AI Security In: MLSecOps in Practice
43mHacking and Securing ML Systems and Environments