Definition
A mathematically proven finding by NIST that no finite set of safety rules applied to an AI system can block every possible harmful input or output—there will always be some edge case an adversary can use to bypass them. This means AI safety guardrails must be treated as a continuously updated defence, not a one-time fix.
Why it matters
Boards and regulators sometimes assume that once an AI system passes a safety evaluation it is permanently safe. NIST's proof shows that assumption is wrong: safety is a continuous process requiring ongoing monitoring, red-teaming, and updates—not a certification that stays valid indefinitely.