Security researcher Ben Zimmermann reports finding 26,400 unique working credentials in 7,256,531 public Gitee repositories. His October 28 account appears on MongoDB’s blog, developed with its Kingfisher team. Gitee is a Git-based source-hosting platform operated by OSChina.

Evidence status: researcher-reported credential exposure, validated at scan time. The article says two scanners authenticated candidate credentials against their providers’ APIs, then deduplicated results by secret value. It supplies no verified count of criminal account takeovers or resulting data theft.

Zimmermann says he notified vendors with the highest exposure volumes. Revocation had not been confirmed at publication. A disclosure email and a revoked key are different outcomes.

An old commit can still hold a working key

The research used TruffleHog and Kingfisher against repository histories. Either tool found credentials the other missed. The reported total includes cloud-provider keys, AI-service tokens and other service credentials, including an organization-administrator credential for an unnamed enterprise AI platform.

A working key can expose resources according to its permissions. That differs from compromising Gitee’s infrastructure or proving that every repository containing a key belongs to an affected employer. The article’s corporate-domain observations come from committer email metadata, which cannot establish credential ownership or organizational impact by themselves.

The month-long scan also limits the meaning of “live.” Authentication succeeded when each key was checked; the article does not establish that every key remained usable on October 28. Nor does its enumerated repository set prove complete platform coverage.

Cover mirrors as well as the main hosting service

For maintainers, the practical inventory includes public mirrors, forks and old repositories, not just the primary GitHub organization. A credential can remain valid after the file disappears from the latest branch.

GitHub’s sensitive-data guidance gives the response order: revoke or rotate an exposed secret first. Rewriting history has coordination costs and cannot retrieve copies held in other people’s clones. Those Git properties apply across hosts, while provider-specific cleanup procedures differ.

For an organization that discovers an exposed key:

  • Disable or replace the credential through its issuer. Update legitimate workloads that depend on it, and confirm the old credential is invalidated. Deleting a repository is insufficient containment.
  • Review the access the credential granted. Establish its scope, lifetime and relevant provider activity records. Successful authentication during research establishes exposure; actual misuse needs separate evidence.
  • Clean up retained copies with collaborators. Include history, mirrors and forks where you control them. Coordinate changes so an old clone cannot accidentally restore the secret.
  • Prevent the next leak. Keep secrets out of tracked files, use narrowly scoped credentials and enable appropriate commit or push checks. Check which credential types a detector supports before treating a clean result as complete coverage.

These are defensive steps for systems and credentials an organization owns. The research is no authorization to test someone else’s exposed keys.

What the thread adds

In the ranked Hacker News discussion, dorianmariecom points out that other Git hosts have the same problem. poly2it replies with the article’s GitLab, Bitbucket and Gitee comparison table. The comparison raises a useful measurement question: were the platforms examined with equivalent detectors?

Zimmermann’s own methodology note answers no. The earlier platform studies used TruffleHog alone, while this Gitee scan used two tools, with different provider coverage. The headline totals therefore cannot rank the platforms’ security on like-for-like terms. Nothing in the replies independently verifies the credentials or establishes their revocation.

From a 4-comment thread, in HN’s own order.

Caveat: MongoDB develops Kingfisher and collaborated on the post; Modal supplied compute-credit sponsorship. The findings and validation claims come from this research account, without independent replication or confirmed remediation.