Breaking Models 🚨AI Security
Details
Breaking Models 🚨 AI Security (Supply Chain Attacks, Part 2)
Last time, we mapped the AI supply chain and broke the model file and the model hub. This time, we finish the map: the steps that change a model, the software under it, the tools your agent loads, and models on your phone. Then you get to play attacker.
Agenda
- Quick recap of Part 1 (new folks welcome)
- Changing models: how quantization can hide a backdoor, and risky adapters
- Pipelines and dependencies: CI hacks, fake packages, exposed AI servers, package names AI made up
- The agent’s runtime: bad MCP servers, poisoned tools, malware that borrows your AI agent
- Models on devices: stealing, fooling and swapping on-device models
- Hands-on: threat model a real-looking AI app in small groups
- Open floor: your questions and your own experiences
Missed Part 1? Come anyway. We start with a 5-minute recap. Built on OWASP GenAI Security Project guidance. Beginners welcome, no ML background needed. No vendor pitches.
📩 Questions, or want to present at a future session? [aisecurityinbay@proton.me](mailto:aisecurityinbay@proton.me)
📍 WHERE Hacker Dojo, 3200 Patrick Henry Drive, Santa Clara.
🚪 HOW TO GET IN Use the front entrance with the automatic doors, then take the stairs or elevator to the 2nd floor Classroom. If the front door is locked, look for a sign on it with directions to the back entrance.
💻 Bring laptops.
