GitHub announced a public preview of secret scanning in Copilot security review on November 17. It is opt-in, checks changes rather than history and consumes AI credits. The listed environments are CLI, the app, SDK integrations and code review; VS Code is not included in this release. This is previously uncovered functionality from the week, not a new announcement today.
The aim is to bring checks closer to the moment when a person or agent proposes code. They can then inform the decision to accept a change. A negative result for one set of changes does not establish that earlier versions contain no secrets. An automated review should not be described as a guarantee of detecting every credential.
GitHub's general documentation separately describes Git-history scanning and recommends rotation after a leak. Deleting the text does not revoke access already granted by a valid key. Preventing a new leak, finding historical traces and invalidating compromised access are therefore complementary tasks.
A team's acceptance question is when a check runs and which revision it covers. If an agent changes more files afterward, the earlier result should not silently apply to those changes. A useful report identifies the checked diff and the technical status of the scan itself. This is a proposed process control.
Distinguishing no findings from an unfinished scan is particularly important. Authentication trouble, interruption or an unavailable service is not evidence that code is safe. A team needs its own rule for who may continue and how that decision is recorded. We do not assume automatic blocking in every preview scenario.
An isolated repository with safe dummy data can be used to examine integration behavior. The purpose is to observe how a result reaches reviewers, not to publish real keys for a demonstration. Even a successful routing test does not evaluate every detector. SIExpo has not conducted such a test.
The new layer can assist early review of agent-written code. Its practical value depends on precise scope, visible failure and a prepared response process. People need to know what was checked, rather than letting a broad feature name create a stronger sense of protection than the evidence supports.



