Author: Nadali. Prepared with AI assistance.
On November 5, 2026, Microsoft described AI agents finding secrets in public GitHub repositories for follow-on attacks. Observations began in August, with subsequent activity from October 28. The news is the research disclosure, not the start of attacks.
Who this helps
Owners of businesses with custom websites, integrations or commissioned software. Using GitHub does not establish that a leak occurred. Ask your contractor about a specific project and accountable person rather than asking generally whether the business is protected from AI.
The practical takeaway is that an exposed secret requires action on access, not just a file edit. GitHub recommends treating a secret committed to a repository as compromised, updating services that use the old token and checking logs. Removing the token from code does not automatically close its alert.
What confirmation to request
An editorial discussion template for the technical owner:
- Which repositories and integrations belong to this project, and who reviews their alerts?
- If exposed access was found, which service is affected, who owns remediation and when will the outcome be checked?
- Which business process needs testing after access replacement: order transfer, stock synchronization or message delivery?
- Where are completed actions, the next check and unresolved questions recorded?
Do not paste keys into the email, task or chat. Use a safe identifier and a link to an appropriately restricted internal task instead.
For illustration, a shop has a stock-transfer integration. After remediation, the responsible person tests an update for one product and records the result. This is an editorial example of a control, not an account of an affected shop.
Limits and the next step
Access-replacement mechanisms depend on the service; its administrator should perform the work and check dependencies. Do not revoke every key indiscriminately because of this report. If a leak is confirmed, do not postpone response until a routine review.
This article does not establish that your business was compromised or promise complete protection. Nadali read the primary sources on November 8, but did not examine the researchers’ raw logs or readers’ systems.
Related reading: keys and sessions after a suspected PC infection.
Cover: conceptual AI illustration, not an incident photograph or a real interface.
