Exchange Online Bug Sends Gmail To Spam: Microsoft Fixes The Issue

An inconspicuous error in a machine learning model from Exchange Online has caused unrest in many mailboxes in the past few days – but now Microsoft has given the all -clear. Microsoft corrects a spam defect alarm in Exchange Online.

Gmail emails were wrongly affected

Since April 25, emails from Gmail accounts have suddenly been incorrectly in the junk folder for numerous users of Exchange online, and not for the first time. Now Microsoft has contained the incident and resolved the underlying problem. The error was triggered by a machine learning model that is part of Exchange’s security architecture online.

This model automatically evaluates incoming messages on potential dangers. But this is exactly where the problem was: the software classified serious Gmail messages as suspicious-apparently because they had similarness to known spam campaigns in structure or content. As a result, the emails were marked as malignant and automatically sorted out.

In the Microsoft-365 administration center, the incident was conducted under the identification ex1064599. An entry of May 1, finally confirmed the defusing when recognizing and sorting. Microsoft reset the machine learning model to an earlier version. Since then, according to the company, no faulty categorization has been determined. Administrators were also able to temporarily emerge with individual filter rules and generate certain senders – such as Gmail.

“Noticeable incident”

It remains unclear how many users were actually affected. So far, Microsoft has not commented on the number of malfunctions or the affected regions. The group only spoke of a “noticeable incident”, which indicates a bigger problem in the internal classification.

Anti-spam rules always faulty

The latest incident is part of a series of similar breakdowns: just last week, another mistake had made an adobe emails to be blocked. In March and last October, incorrect anti-spam rules had to be withdrawn or deactivated. Even a simple image attachment was able to lead in August 2024 to be classified as a threat and put under quarantine.

Microsoft emphasizes that one is continuously working on the refinement of ML detection. The aim is to trigger less false alarms in the future without weakening the protective effect. How reliable mechanical learning can be in security -critical areas is repeatedly put to the test.