Nondeterministic software can produce different results from the same input because its behavior depends on hidden state, timing, randomization, external data, or model inference. In AI systems, the same prompt may lead to different responses, tool choices, or refusal decisions even when the visible input has not changed.
This matters in cyber security because unpredictability makes systems harder to test, audit, and defend. A security control that works once may fail on a later run, which complicates incident reproduction and increases the chance that unsafe output slips through. Attackers can exploit this by repeatedly submitting the same malicious content until the system takes a dangerous action, while defenders reduce risk by limiting privileges, isolating untrusted inputs, validating outputs, and requiring human confirmation for sensitive actions.