AI Safety Forum workshops
Keynotes and five hands-on workshops on how modern AI systems fail, and how we find out before they are deployed. Everything runs in the browser; the heavier experiments run on a shared backend, or on an H100 if you have an access code.
Have an access code? Use the “Unlock the H100 backend” box under the header. It is also available on every page.
Keynotes
Slides, PDF downloads and presenter notes for each talk. Decks that belong to a workshop link to it.
Interactive workshops
Bias–Variance Tradeoff
Train polynomial house-price models, play against your model, and see over- and underfitting with your own eyes.
Fundamentals · Module 2Robustness & Adversarial Attacks
Train a traffic-sign classifier on GTSRB, evaluate it, then fool it with patches, noise, FGSM and one-pixel attacks.
Generative AI · Module 3Red Teaming LLMs
A group exercise probing language-model safety: personas, refusal suppression, obfuscation and more.
Evaluation · Module 4.4AI Benchmarking and How It Breaks
Can a model tell it is being tested? A quiz and a small reward-hacking game.
Generative AIPoisoning Training Data
Audit SecurePay's payment-risk model, then find the backdoor trigger an adversary planted.
How to use these pages. Each workshop opens with a link to the matching keynote slides. Work through the steps top to bottom; the controls let you try things out rather than follow a fixed path. The material presents options and findings, it does not prescribe a single right answer.