A Java 21 Spring Boot image on Debian 12 started at 839 MB with 101 CVEs and 244 packages. After curating and then optimizing it, the same application ran on a 241 MB image with zero CVEs. That result took both steps: curating patched the image, and optimizing removed what the application didn't use. The difference between curated vs hardened container images decides how many CVEs you carry, how large your images are, and how much Dockerfile rework your engineers will accept.
This article defines each type, shows how to adopt a curated image without breaking your build, explains how compliance requirements narrow the OS choice, and walks through starting points for new and existing applications. The content of this article is based on a webinar titled “Curated images: the foundation of near-zero CVE software supply chain”.
Curated vs hardened container images: what each one changes
A curated image is a patched copy of an open source image. The patches are backported, so the image stays on the same OS and version as the one you use today. A production-grade curated image is tested and verified to function, built on long-term support (LTS) releases, rebuilt with new patches once or twice a day, and ABI-compatible (Application Binary Interface compatible) with the original, so it works as a drop-in replacement. Curating doesn't change what the image does or, in like-for-like comparisons, how large it is.
A hardened image has had components removed: the package manager, shells, and other utilities that belong in development, not production. Hardening on its own patches nothing. In the webinar, hardened images were about 75% smaller than the originals.
A curated and hardened image is both patched and minimized. RapidFort calls the minimizing step optimizing, and the result an optimized image. It gives you the fewest CVEs in the smallest footprint.
| Curated | Hardened | Curated and hardened | |
|---|---|---|---|
| What changes | Patches applied and backported; no functional change | Unneeded components removed; no patching | Patched and minimized |
| Image size | About the same as the original | About 75% smaller | Smallest footprint |
| Optimized for | Fewest CVEs | Smallest size | Fewest CVEs and smallest size |
| Best used as | Golden base images for developers | Minimal runtime images | Production images |
Compare sizes like for like. A curated distroless image should be about the size of the original distroless image, and a curated full-size OS should be about the size of the original full-size OS.
Why drop-in compatibility decides adoption
Engineers adopt a curated image only if it doesn't break anything, and that depends on matching the OS and the version exactly. Most Dockerfiles are written in one distribution's package manager syntax. Moving from Ubuntu to Red Hat, or from Red Hat to Alpine, means rewriting those Dockerfiles and retesting drivers and dependencies. A robotics application that depends on Ubuntu-specific drivers, for example, turns a base image change into a migration project.
Versions work the same way. Upgrading from Python 3.13 to 3.14 can pick up a fix, but many teams won't take the upgrade until they know 3.14 is backward compatible with their code. A curated Python 3.13 image on the same Debian base keeps the major version, minor version, and OS identical, so only the patches change.
In practice, the switch is one line in your Dockerfile:
# Before FROM python:3.13-[INSERT ORIGINAL TAG] # After FROM [INSERT RAPIDFORT REGISTRY PATH]/python:3.13-[INSERT CURATED TAG]
Because the OS, packages, and source code don't change, your standard QA or smoke test is enough. If the OS changes, you need full application testing, because you can't predict how the new OS and its drivers will behave.
You can keep pinning images for reproducible builds. Tracking the latest tag picks up each day's patched build. Pinning to a version, such as Python 3.13, picks up the newest patched build of that version. Pinning to a digest keeps exactly that image until you change the pin.
First-party and third-party images follow different rules
First-party images are the ones your team builds on. You pick an OS and a framework, then write your application on top. Your Dockerfile names the OS and its packages, so you need a curated image that matches them exactly.
Third-party images are complete functions you run as they are, such as Redis as a database or NGINX as a load balancer. You aren't editing their Dockerfiles, so the underlying OS matters less, and you can choose the OS that fits each workload.
For a load balancer, an Alpine-based image gives you the smallest footprint. Alpine has no Security Technical Implementation Guide (STIG), so if you're working toward FedRAMP, an Ubuntu or Red Hat variant, both of which have a STIG, is the better fit. Swapping the OS on third-party images lets you trade footprint against compliance one workload at a time.
How compliance requirements narrow the OS choice
Start from the end state: the compliance test the image has to pass.
STIG: FedRAMP requires a STIG approved by the Defense Information Systems Agency (DISA). According to the webinar, the Linux distributions with DISA-approved STIGs today are Ubuntu, Red Hat, Amazon Linux, Solaris, SUSE Linux, and Oracle Linux. Some minimal distributions publish their own benchmarks, but DISA hasn't validated them, so using one means requesting a deviation, which slows your authorization.
CIS Benchmark: Where a STIG isn't required, a CIS Benchmark can be enough. Some distributions, Alpine among them, have neither a STIG nor a CIS Benchmark.
FIPS: For FedRAMP or an authority to operate (ATO), choose a variant with FIPS-validated cryptography and confirm that it uses FIPS 140-3 modules.
Requirements differ across FedRAMP, CMMC, SOC 2, and the EU Cyber Resilience Act (CRA). FedRAMP needs a STIG, while CMMC or SOC 2 can be satisfied with a CIS Benchmark. Confirm the framework before you choose the image.
Pick a starting point for green field and brown field applications
The right base depends on whether the application is new or existing, and on how much container expertise your team has.
| Starting point | Approximate size | Best for |
|---|---|---|
| Full-size curated OS (for example, a curated FIPS Ubuntu) | 80–85 MB | Legacy applications where the original team has left and nobody can say what is safe to remove |
| Smaller curated OS with a package manager | 50–60 MB | Teams that want to choose which packages to add |
| Distroless | About 10 MB | New builds |
| Scratch | About 500 KB | New builds by experienced teams using multi-stage builds |
A scratch image contains no shell, package manager, or utilities. Your team adds everything deliberately, line by line, so the image stays small, which suits green field work by experienced builders. For brown field work, swap the existing OS for an identical curated version first and optimize at the end.
Curate first, then optimize
Start with a curated image at whatever size fits your team and add the components your application needs. Then profile the image at runtime to produce a Runtime Bill of Materials (RBOM®) that shows which packages and files the application uses. RapidFort removes the unused ones, down to the file level. Run the optimized and unoptimized versions side by side in QA so you can compare them before you switch.
Two examples from the webinar show the effect:
| Image | Original | After curating | After curating and optimizing |
|---|---|---|---|
| Python 3.13 on Debian 12 | 1.14 GB, 167 CVEs | 250 MB, 4 CVEs | 89.2 MB, 3 CVEs |
| Java 21 Spring Boot on Debian 12 | 839 MB, 101 CVEs, 244 packages | 528 MB, 4 CVEs | 241 MB, zero CVEs |
A minimal starting image gets you only part of the way. Teams keep adding packages, AI coding assistants pull in more dependencies, and nobody goes back to remove what is unused. Curating and then optimizing is how RapidFort customers reach up to 99.9% CVE elimination and up to 90% attack surface reduction.
Results depend on the language. The webinar described more than 90% attack surface reduction for Python, Node, and Java images, and 30% to 35% for compiled languages such as Go and Rust or for images that started distroless, because there is less left to remove.
Match the image to the workload
Use curated images as your developers' golden base images and curated and optimized images in production. Match the OS and version exactly so adoption is a tag change, let compliance requirements set the OS for third-party images, and pick a starting size that fits your team and application. For background on the optimizing step, read What is container hardening?.
Swap your current base image with a one-line change. RapidFort Images cover the major LTS distributions, with curated and optimized variants for production. Learn more about RapidFort Images
Latest posts
Eliminate Attack Vectors at the Source
Continuously eliminate up to 99.9% of CVEs without code changes
Products
Use Case
Quicklinks
Stay in touch
Subscribe for product updates and RapidFort newsletter.
Address
440 North Wolfe Road, Sunnyvale, CA 94085
© 2026 RapidFort, Inc.
RapidFort, RAPIDFORT, and RBOM® are registered trademarks of RapidFort, Inc. All other marks and names mentioned herein may be trademarks of their respective companies.






