DVC 資產盤點應追蹤遠端儲存及憑證檔案

DVC inventory should follow remotes and credential files

Data Version Control, or DVC, manages data and model files through Git-like metadata while storing the actual content in remotes such as S3, Azure Blob Storage, Google Cloud Storage, SSH, SFTP, HTTP, WebDAV, HDFS, or local directories. A DVC application fingerprint is therefore not a reliable proxy for the data exposure. The important assets are the configured remotes and the credentials used to reach them.

Measurement

On 2026-11-03 UTC, ZoomEye was queried with page 1, pagesize=1, and sub_type=all. app="DVC" returned 16 observations. title="DVC" returned 802.

The application count is a small set of observable matches. The title count is broad and may include documentation, tutorials, and unrelated uses of the acronym. Neither number should be presented as the number of exposed data stores.

The remote is the trust boundary

DVC's documentation says remotes are external storage locations used to track and share data and models. Authentication and other remote settings can be configured with dvc remote modify. Sensitive values should use --local so they are written to .dvc/config.local, which is typically Git-ignored and not shared with the repository.

That local configuration is the artifact to protect. A public repository can safely contain the remote URL while the credential path, password, token, or private-key setting stays local. If an operator puts a secret in the shared .dvc/config, every clone may inherit it. If a build job writes the secret into an artifact or log, the remote may be reachable even when the DVC application is not public.

Inventory the data path

Begin with the 16 application observations and internal repositories that use DVC. Identify each remote type, URL, credential source, and data classification. Check whether the remote is public, bucket-policy protected, SSH-key protected, or behind an internal gateway.

Review local configuration files separately from source-controlled files. Search CI/CD definitions for commands that materialize credentials from environment variables, then confirm that those variables are not printed in logs. For S3 and other cloud remotes, review the remote's own IAM policy and access logging. A DVC-level configuration cannot compensate for a public bucket.

Treat the counts as leads

The 16 app matches may include services, dashboards, demos, or unrelated fingerprints. Verify ownership before taking action. The 802 title matches should be sampled for clear product pages and used as a discovery aid, not counted as data repositories.

For each confirmed DVC project, test that an unauthorized user cannot clone the private configuration or read the data remote. In a controlled environment, attempt a dvc pull without the expected credential and confirm that the remote rejects it. Review whether the ignored local configuration is included in backups, container images, or support bundles.

Practical response

Keep credentials out of the repository and DVC-tracked content. Use cloud-native temporary credentials or scoped service accounts where possible. Restrict the remote to the network and identities that need it, and enable access logging on the storage side.

If a credential is exposed, rotate it and review remote access logs before deleting evidence. The relevant question is which objects were read, listed, or changed, not only whether the DVC client was installed. If the remote permits public listing, fix that policy first.

The measurement lesson

DVC makes the data path visible through configuration. The security team should follow that path. A count of 16 application matches can help find deployments, but the real inventory is a list of repositories, remotes, credential sources, data classifications, and verified access policies.