I'm not sure that's actually true in the case of ThreadSanitizer - a well-designed data race detector can detect race conditions that didn't actually get triggered in that run. In any case, usually people try and exercise those rare code paths by running a fuzzer against sanitizer-instrumented code.
ThreadSanitizer can also only catch bugs that can occur in the code paths that are exercised – and running a fuzzer can certainly show the presence of bugs, but I'd be wary if someone claimed it can show the absence of them.
