What happened?
Microsoft's AI research team posted an open-source code-training dataset on GitHub (the robust-models-transfer repository) and used an Azure SAS token link to share files. The link was configured to share the entire storage account instead of only the necessary files, with full control permission instead of read-only, and an expiry stretching to 2051.
The cloud security firm Wiz discovered the incident. According to Wiz's report, the account held an additional 38 TB of data, including disk backups of two employee workstations: Microsoft service passwords, secret keys and more than 30,000 internal Microsoft Teams messages from 359 employees. The link had been misconfigured since 2020.
Wiz notified Microsoft in June 2023; the token was disabled soon after and the incident was made public on 18/9/2023. Microsoft said no customer data was exposed and no other internal services were put at risk.
Why is it serious?
- Write permission would have let an attacker modify the AI models in the repository, so anyone trusting Microsoft's GitHub repo would have downloaded models injected with malicious code.
- One simple sharing action pulled in entirely unrelated data.
- SAS-style links are hard to monitor centrally; token-creation mistakes often go undetected for years.
Lessons for enterprises
- Share with the right scope and the right duration. Every sharing link needs limits on files, permissions and an expiry date.
- Backups are sensitive data too. Workstation backups contain passwords and keys, so they need encryption and permissions just like the original data.
- Periodically review open links. Without a review, you do not know what you have shared.
A data-centric approach
When policy travels with each file (encryption, permissions, expiry, revocation), a link or a backup slipping out does not mean the recipient can open and use the content. Enterprises set an expiry for shared documents, revoke rights when no longer needed and read access logs to know who has touched the data. See DCS features and the enterprise assets risk group.
Reference sources
- Wiz Research: 38TB of data accidentally exposed by Microsoft AI researchers
- InfoQ: Microsoft AI researchers accidentally exposed 38TB of sensitive data
