AI Goat: A Damn Vulnerable AI Infrastructure
summary
Deliberately vulnerable lab called AI Goat reproduces OWASP AI Top 10 risks and yields three hands‑on challenges; deployment uses Terraform and Github Actions to spin up an AWS stack with models on AWS SageMaker and a web frontend, and it actually retrains when a publicly writable product-ratings CSV is altered so recommendations flip. Demos show an image‑metadata command execution, data poisoning via writable bucket, and an output‑integrity bypass by forging model prediction fields — all reproducible and open source.
AI-written summary · about
chapters (9)
- 0:00AI Goat introduction
- 4:58Securing AI infrastructure
- 6:23What is AI Goat
- 12:52AI supply chain attack demo
- 18:49Data poisoning attack demo
- 26:40Output integrity attack demo
- 32:00AI environment protections
- 35:09AI Goat future plans
- 36:14Q&A
related talks
57mAI Applications Security Puzzle
44mAI Showdown: The Good, The Glitchy, and The Risky
17mAI Safety in the Generative AI Era
38mA tale of 3 goats - Immutable infrastructure security
36mThe Dark Side of AI: The Hidden Risks in Open-Source AI Models
25mPreventing an OWASP Top 10 in the world of AI
52mAI Is Just Software, What Could Possibly Go Wrong?
56mHidden Risks in Open-Source Code and AI Models