Skip to main content

What DockSec does not do

Stated plainly, because this is what you need before letting a tool gate your builds. Every item here is a real limit, not a roadmap tease.

Reachability is not proven​

A finding means the vulnerable version is present in the image. It does not mean the vulnerable code path is reachable, or that your application calls it. DockSec says so in its coverage notes on every run.

EPSS narrows the list to what is being exploited somewhere. It does not tell you whether it is exploitable in your image.

Chains cover Compose only​

Exploit chain detection reasons over a Docker Compose topology. It does not cover Kubernetes manifests, Helm charts, or chains spanning a Dockerfile and a Compose file. A single-service Compose file has no cross-service chain to find.

Offline mode has a precondition​

--offline uses the Trivy database already on disk and skips the update. There is no bundled advisory database, so an air-gapped machine needs a Trivy DB copied to it first. "Air-gapped" is supported with that condition, not unconditionally.

The AI pass is verified against one provider​

The correlation pass has been tested against Claude. OpenAI, Gemini and Ollama are wired up but their output quality is unverified. The deterministic core - everything under --scan-only - does not use a model at all and is unaffected.

There is also no live-model test in CI. That path is covered by a stubbed client, which by construction cannot catch a malformed real response.

Severity mapping for Hadolint is a mapping​

Hadolint reports error / warning / info / style. DockSec maps those to HIGH / MEDIUM / LOW / LOW. That is a deliberate, conservative convention, not a measurement, and reasonable people would map it differently.

SBOM is CycloneDX only​

SPDX is not emitted.

The score is exposure, not a grade​

The score is deterministic and severity-weighted. It does not discount a finding for being unfixable, because "you cannot fix it" and "it does not matter" are different statements. A well-maintained image with no available patches can still score poorly - see the python:3.12-slim case study.

A clean result is not a guarantee​

Zero findings means no known, published advisory matches a package in that image today. Tomorrow's advisory applies to the same bytes. Re-scan on a schedule and use --baseline so CI tells you when a previously clean image stops being clean.