We've noticed occasional errors in our system that we were unable to reproduce reliably and couldn't track down easily (because of a unrelated) bug in some Sun code. A coworker has finally found the cause and produced the following analysis: This method of org.codehaus.backport175.reader.bytecode.AnnotationReader public static AnnotationReader getReaderFor(final ClassKey classKey) { final AnnotationReader reader; Object value = READERS.get(classKey); if (value == null) { synchronized (READERS) { reader = new AnnotationReader(classKey); READERS.put(classKey, new WeakReference(reader));//reader strong refs its own key in the weakhahsmap.. } } else { reader = (AnnotationReader) ((Reference)value).get(); } return reader; } contains an error. The AnnotationReader that is obtained in the line reader = (AnnotationReader) ((Reference)value).get(); can be null, as the WeakReference content can be nullified by the VM if no other references to the content exist. In practice, this happens very seldom and non-deterministically. But as it sometimes does occur, null is returned in this case, which leads to a NullPointerException where the reader is used (in the Annotations class). A proposed fix is public static AnnotationReader getReaderFor(final ClassKey classKey) { AnnotationReader reader; Object value = READERS.get(classKey); if (value == null) { synchronized (READERS) { reader = new AnnotationReader(classKey); READERS.put(classKey, new WeakReference(reader));//reader strong refs its own key in the weakhahsmap.. } } else { reader = (AnnotationReader) ((Reference)value).get(); if (reader == null) {//WeakReference content can be null synchronized (READERS) { reader = new AnnotationReader(classKey); READERS.put(classKey, new WeakReference(reader)); } } } return reader; }