Threat Modelling for ML/AI systems
summary
Two-hour workshops adapt threat modeling for ML/AI by splitting training vs inference, running on Azure, and producing concrete fixes: disable Microsoft abuse monitoring for an internal ChatGPT instance, implement rate limiting on seismic jobs, tighten admin access and MFA, whitelist frontend links, and consider model encryption. Teams run 2‑hour sessions with online whiteboards, find about 30 threats identified across four cases, emphasize training data manipulation as most frequent, use traditional app‑security controls, and involve seven colleagues in the enabling team.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:25Context
- 2:01Security Champions
- 8:46When to apply Threat Modelling
- 10:17How to apply Threat Modelling
- 11:27Identifying Concerns
- 13:32Adapting to MLAI
- 14:25Training vs Inference
- 18:14Input Data Manipulation
- 23:53Actions
- 25:43Automatic seismic interpretation
- 30:13MLOPS platform
- 33:17Trading recommendation system
- 36:51Summary
related talks
45mThreat Modeling in the Age of AI
25mThreat Modeling for AI Apps with Attacks as Code
29mThreat Modeling Meets Model Training: Web App Security Skills...
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
45mHardening AI/ML Systems - The Next Frontier of Cybersecurity
44mThreat Modeling to Protect Your Generative AI Applications
48mTeaching Software Engineers to Threat Model: We Did It, and So Can You
45mHow Three Threat Modelers Can Influence an Entire Organization