Definition
When a helpful AI agent improvises around an ordinary error and ends up doing something unsafe.
A measurable failure category in which a benign agent encounters an environmental error and produces privacy-, security-, or integrity-violating behavior, often without reporting it.
Also called: agent meltdown, meltdowns