GitHub moved secret scanning to a purpose-built model: what changes for hardcoded passwords
Hardcoded passwords are the finding every secret scanner is worst at. An AWS key has a prefix and a fixed shape. password = "hunter2" has neither, so a pattern match either misses it or fires on every test fixture in the repo.
On October 7, 2026, GitHub announced a fine-tuned model for leaked secret detection aimed at exactly that case. In the changelog's words it "reads surrounding code to identify likely credentials, including passwords without a recognizable token format." It does not generate code or prose. It sits behind three things: secret scanning alerts, push protection, and Copilot security reviews.
What is on, and what is not
As of October 15, 2026, going by the changelog:
| Piece | State | Cost |
|---|---|---|
| AI-detected Password alerts | Customers were upgraded to the new model automatically | Included with GitHub Secret Protection and GitHub Advanced Security |
| AI push protection | Opt-in; an administrator enables it | Uses GitHub AI Credits |
Copilot /security-review secret checks | Coming soon in private preview, off by default | Uses AI Credits, billed to the Copilot plan |
| GitHub Enterprise Server 3.23 | AI-detected alerts in public preview | No push protection, no security review |
One inconsistency to know about: the page calls AI push protection a private preview in the summary and a public preview in the billing section. Treat it as a preview and ask your account contact.
The false positive question
Does the new model produce fewer false alerts? The changelog does not say. It publishes no precision figure, no false positive rate and no list of languages or file types. The stated idea, reading the code around a string instead of matching its shape, is plausible. A plausible idea is not a measurement.
Nor does the page say what stays with patterns. Its own description covers "passwords without a recognizable token format" and unstructured credentials at push time. Provider tokens with a fixed prefix and shape are the case where a pattern is exact and cheap, and the changelog never claims the model replaces that. Read it as an addition for the shapeless leftovers, not a swap.
So measure it. Take two weeks of the new Password alerts and mark every closed one as real, test data, or placeholder. That count tells you more than any announcement, and it is the same exercise as asking how many of your alerts anyone acts on. If most of them are placeholders, you have a new noise source with a better name.
The billing trap
The free part is the alert scan. The opt-in parts draw AI Credits, and the page says usage begins "in the coming weeks". Push protection usage is billed to the organization that owns the repository, even when the check does not block the push.
The changelog says it plainly: "Budget alerts alone don't stop usage." Its advice is to create a SKU-level budget with Advanced Security as the product and Secret Protection AI Credits as the SKU, and to turn on "Stop usage when budget limit is reached" where that option is available. The exact menu path is in GitHub's budget documentation, so follow that rather than my summary.
Teams already on the AI push protection preview are warned that continued use after the billing change will consume AI Credits. If you do not want that, turn it off first.
No per-credit price is on the page, so check the billing documentation before enabling it org-wide.