Pimp My IDE / garage dispatchBack to the garage
October 8, 2026 / secret scanning / cost control

A better detector still needs a fuse.

GitHub's new classifier can catch credentials that do not look like tokens. Detection, push blocking, billing, and credential rotation remain separate jobs.

Wire the catch. Name the consequence.
SECRET PROTECTION / CIRCUIT MAPNO LIVE SCAN
DETECTFind likely secret
BLOCKStop the push
BILLCap AI credits
ROTATEKill exposure

A classifier verdict is an alert. It is not proof that the credential is active, revoked, removed from history, or safe to ignore.

The detector reads context now.

GitHub announced a purpose-built model that reads surrounding code to find likely credentials, including passwords without a recognizable token format. The company says existing AI-detected password alerts have moved to the new model. AI checks for push protection are in private preview, and checks inside Copilot's /security-review command are planned for private preview.[1]

This fills a real gap. A regular expression can match a known token shape. A password in a configuration assignment may have no stable prefix, length, or alphabet. Context gives the detector another signal.

A finding can stop a bad push. It cannot rotate the credential that already escaped.

Three surfaces, three contracts.

Repository alert scanning, push protection, and a local security review run at different times. An alert can arrive after content enters repository history. Push protection can interrupt the write before that happens. A security review can inspect active changes before commit or push.[1]

Do not reduce these surfaces to one green shield. Record where the check runs, what it can block, who can bypass it, and who receives the finding. The new push and security-review checks are opt-in. Running /security-review alone will not enable the new classifier.[1]

The cost switch is part of security.

GitHub says AI-detected alerts remain included with Secret Protection. The planned push-protection and security-review checks will consume AI Credits after customers opt into the preview and enable them. GitHub also says a push check can consume credits even when it does not block the push.[1]

A budget alert is not a hard stop. GitHub tells administrators to configure "Stop usage when budget limit is reached" when they need a spending cap.[1] That setting belongs beside the enable switch. An agent must not change either without explicit authority.

Rotation closes the incident.

GitHub's secret-scanning guide says a committed credential should be treated as compromised. The response can include validity review, token deletion and replacement, service updates, and a check for unauthorized activity.[3]

Removing the string from the branch is not the same as revoking it. GitHub also notes that removing a token does not automatically close the alert. The alert needs a separate resolution step.[3] Keep detection, code cleanup, rotation, and alert closure as distinct receipt fields.

Interactive makeover / scanner toggle plus incident fuseboard

Wire the secret circuit.

This replaces one vague "AI scanning on" switch with a mode selector, four independent review fuses, and a copyable rollout card. It is better because the selected surface changes the documented block point and payer while the response gates stay visible. This teaching component does not scan code or change GitHub settings.

DetectBlock pointCost ownerRotate

Surface selector

Pick where the check runs. Then select the review fields your rollout card should request.

Detection surface
Rollout fuses
Rollout card / structure only

Secret circuit receipt

REPOSITORY ALERTS / AFTER ENTRYFindings can flag repository content. They do not block the original write.
0 rollout fuses selectedSelect the fields this card should request. No repository, policy, budget, or rotation evidence is attached.
Sources read / October 8, 2026

Source log

Open the evidence boundary
  1. GitHub Changelog, "Purpose-built model for leaked secret detection", published October 7 and read October 8, 2026. GitHub describes the classifier, the three product surfaces, preview status, plan coverage, AI Credit rules, opt-in behavior, policy controls, and budget-stop guidance.
  2. GitHub Docs, "Secret scanning", read October 8, 2026. GitHub documents the repository surfaces scanned, generic and custom patterns, validity checks, AI-detected secrets, plan coverage, and public monitoring.
  3. GitHub Docs, "Resolving alerts from secret scanning", read October 8, 2026. GitHub says to treat committed secrets as compromised, rotate affected credentials, review service use and logs, and close alerts separately after remediation.
  4. GitHub Changelog, "Local sandboxing for GitHub Copilot now generally available", published October 7 and read October 8, 2026. GitHub states that model execution and tool isolation are separate concerns. This supports the same separation between model-based detection and the controls around agent actions.

Evidence boundary: we read GitHub's announcement and documentation. We did not enable the previews, spend AI Credits, scan a repository, test detector accuracy, or handle a live credential. The fuseboard drafts a review request. It does not report a security verdict.