Why Dependency-Track, Syft and Grype

Syft, Grype and Dependency-Track are the lab’s software supply-chain layer. They exist because a registry comparison answered a question the human had not asked, and the human noticed.

Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by the site owner. How this site is written →

The question that was answered instead

The human’s stated goal was a reasonably secure supply chain, with tracking, and alerting and possibly remediation if a compromised package ever landed in the registry or was deployed in the environment. The agent had compared Gitea against Zot, Devpi, Verdaccio, Harbor and Nexus. The human asked whether the agent had actually compared the tools in the runbook against Gitea, and it had not.

All six of those products are registries. They answer where approved packages live and how CI gets them without the internet. None of them answers what is inside an artifact, whether any of it is known to be bad, or whether that changes later. The agent said so plainly in the record: the earlier decision still held for storage, and it was irrelevant to the goal.

LayerToolsQuestion it answers
StorageGitea, and the registries it was compared withWhere do approved packages live
AssuranceSyft, Grype, Dependency-TrackWhat is inside them, is any of it known-bad, and tell me when that changes

Harbor would not have delivered the goal either. Its scanning is point-in-time, at push. “If a compromised package ever lands” is continuous by nature, and a build-time scan only knows what was public on build day.

A trap in the offline requirement

Both Grype and Dependency-Track need vulnerability data, and both normally update it from the internet. The lab’s other constraint is that CI must not fetch from the internet. These are compatible only if designed together: Grype can run from a pre-staged database with auto-update off, and Dependency-Track can read a mirrored feed.

If that is left undesigned, the first pipeline either reaches out, defeating the constraint, or runs against a stale or empty database and reports no vulnerabilities on an artifact it never scanned. That is the worst failure available, a green tick that means nothing, and the proof required is that the scanner is made to fail on a deliberately vulnerable fixture before its passes count.

Staged, cheapest first

The recommendation did not start with the heavy service. Syft and Grype go in the CI image and every pipeline run emits a software bill of materials and a scan report, with no new host. The bills of materials are kept, as packages keyed by artifact digest, which is nearly free and gives the later stage a history to start from. Dependency-Track, the piece that satisfies the continuous half, comes third, because standing it up first would mean a heavy service with nothing to ingest. Extending coverage from CI artifacts to deployed services, one bill of materials per running service, comes last and is the largest piece.

Human’s call

Two decisions the record could not settle. Dependency-Track is a JVM wanting several gigabytes, and the proposed host had little margin, no swap and also ran the git origin, so an out-of-memory kill might not choose Dependency-Track. The human chose to place it elsewhere, in its own container.

And the human ruled out automatic remediation: alert, keep the inventory, and give the human automation to use if they choose to remediate. An automatic rebuild-and-redeploy triggered by a feed is a supply-chain attack surface of its own.

What it buys, stated honestly

It detects known-vulnerable components, which is most of real risk. It does not detect a fresh malicious release. A compromise often has no CVE for days, and the incident that led the runbook to replace its earlier scanner was exactly that shape: the scanner would have said clean the morning it happened.

What protects against that is pinning by digest, so nothing new is pulled by accident, ingesting deliberately, so something enters only when someone chose it, and holding an inventory, so that when a disclosure lands the question “am I affected” takes minutes. The realistic value is inventory and speed of answer, not prevention, and it is still the difference between hours and weeks.

What it costs

A heavy service to keep running and backed up, a vulnerability feed to mirror, and a scan whose cleanliness only means something while the database behind it is current.