Adversarial Engineering

Stop Trying to Teach AI Values. Use Type Systems to Lock It Down.

The AI industry is obsessed with ‘goal alignment’—hoping to teach machines human values. But relying on probabilistic models to internalize ethics is a dangerous bet. Martin Odersky’s award-winning research proposes a better way: tracking capabilities in type systems to enforce architectural constraints, making agents safe by locking down what they can physically do.

The Plane That Lands Itself Is the Most Dangerous Thing in the Sky

Emergency auto-land tech saves pilots from heart attacks, but it also degrades their manual flying skills. The real danger isn’t the moment of crisis—it’s the slow erosion of human capability that makes every flight dependent on software. The safest pilot is the one who never has to fly, until the code fails.