Anomaly-based intrusion detection promises to detect novel or unknown attacks on industrial control systems by modeling expected system behavior and raising corresponding alarms for any deviations. As manually creating these behavioral models is tedious and error-prone, research focuses on machine learning to train them automatically, achieving detection rates upwards of 99 %. However, these approaches are typically trained not only on benign traffic but also on attacks and then evaluated against the same type of attack used for training. Hence, their actual, real-world performance on unknown (not trained on) attacks remains unclear. In turn, the reported near-perfect detection rates of machine learning-based intrusion detection might create a false sense of security. To assess this situation and clarify the real potential of machine learning-based industrial intrusion detection, we develop an evaluation methodology and examine multiple approaches from literature for their performance on unknown attacks (excluded from training). Our results highlight an ineffectiveness in detecting unknown attacks, with detection rates dropping to between 3.2 % and 14.7 % for some types of attacks. Moving forward, we derive recommendations for further research on machine learning-based approaches to ensure clarity on their ability to detect unknown attacks.