The paper studies the early detection problem of stability losses or close-to-instability conditions in feedback-control systems where the plant dynamics are uncertain and possibly time varying. A set-point control environment is considered where normal operating conditions imply the presence of a persistent piece-wise stationary noise with possibly time-varying intensity, impulse and step-wise disturbances and no excitation by the reference over long time intervals. The study unveils the existence of a scalar statistic (residual) which can be used to build the desired detection test under the stated prescriptions. The test hinges upon a small-gain stability lemma which ensures that, under divergence, the statistic certainly exceeds a precomputable level. An example illustrates the use of the test within the context of self-scheduling switching supervisory control. (C) 2003 Elsevier Ltd. All rights reserved.