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

[Shmi]

That’s not a bad framing, wonder if it can be formalized.