% █ CyberAI dinner - Building Security Gyms for Eval and Training
Special Lead: Samuel Amouyal - Research Scientist @ Legion Security. Sam is going to share about their recent research building a simulation engine for defensive security.
Simulations mean different things to different people — red teams use them to test controls, AI teams use them to train agents, detection engineers use them to regression-test rules. Same word, completely different gyms.
Let's compare notes on the why, what, and what you actually got out of it:
- Why did you build yours — control validation, agent training, detection regression, something else?
- What's in your stack: live execution, frozen captures, synthetic logs — and why that choice?
- Getting ground truth right — harder than the attacks, rarely talked about
- What signals told you the simulation was realistic enough to trust?
- The gap between "our eval looks great" and "our agent flopped on a real incident"
Bring examples — what you built, what surprised you, what you'd do differently. Or prepare by doing some research - notebookLM with content.
Samuel Amouyal paper:
Don't forget: Bring your own dinner – cook, Wolt, whatever.
Same crew, same vibes 😊