tailcalled comments on Don’t design agents which exploit adversarial inputs

[tailcalled]

Maybe if you have a good measure of being in-distribution, which itself is a nontrivial problem.