Hack your way into LLMs
Can you read a model's mind before it writes the exploit?
SAIL is hosting a mechanistic interpretability hackathon at EPFL. A curated adversarial dataset. Five coding models. One question: what's actually happening inside when they decide to help an attacker?
Two tracks:
Probes as monitors — classify internal activations to detect harmful compliance before generation lands
Attribution — pinpoint which inputs actually push a model toward an attacker
What you get: Claude Code credits, GPU instances, mentors, and food all weekend. A real shot at a publishable result in 48 hours.
Prizes: CHF 1,000 · CHF 500 · CHF 300
PyTorch fluency + curiosity about model internals. That's the bar.
Register by May 6th
May 9–10 · Teams of 5 · EPFL Hosted by SAIL × MLO Lab.