Skip to content

A ContainerOps Exposure Audit: Ports, Users and Limits Checklist

Published November 16, 2026 by Beena Yadav

Set aside one afternoon and pick a single machine that hosts containers. The audit below needs no new software, just a terminal, a notebook and a habit of recording what the host really does rather than what the team remembers configuring. The structure follows the territory the Ops Atlas assigns to ContainerOps: the care of individual containers, their images and their immediate runtime settings, with orchestration and deep supply-chain work left to other disciplines.

Take the sections in sequence. Against every line, note pass, fail or unknown; an unknown stays a failure until somebody produces evidence. The reference for the practices is the ContainerOps entry in the Ops Atlas, which is worth keeping open in another tab.

Section 1: What is really listening on the network

Start here, because the surprises are largest. The Atlas stresses that ports published by the container engine bypass host firewalls such as ufw and firewalld, since the engine writes its own forwarding rules into the NAT table ahead of the point where the firewall looks. A rule that says deny can therefore be quietly skipped.

  • List every published port on the host and match it to a service you can name.
  • For each, read the address it is bound to. A mapping written without a host address binds on all IPv4 and IPv6 interfaces, which Docker’s own documentation describes as insecure by default.
  • Rewrite internal-only mappings to bind to 127.0.0.1, then reach them through a reverse proxy or an SSH tunnel.
  • For services that only talk to each other, move them to a network marked internal so no outside traffic can reach them at all.
  • From a second machine, scan the host’s public address and compare the open ports with your list.

The last item is the one people skip. Believing the firewall is configured is not the same as checking what an outsider can reach.

Section 2: Who the processes run as

  • Inspect each container and record its user. Anything running as the administrator account is a finding.
  • Check whether the container drops capabilities. Its standard advice is to drop every one and add back only the few the program proves it needs.
  • Check whether the root filesystem is read-only. Where it is not, ask why the application needs to write there.
  • Search for the privileged flag in every run command and Compose file. Each hit needs a written justification or a fix.
  • Search for mounts of the engine’s socket. In the Atlas’s reading, that access and the docker group both amount to root on the host.

For anyone moving from a sysadmin background, this section will feel familiar: it is the least-privilege review you already know, applied to a new kind of process. The path described in our walk-through for support and sysadmin staff heading toward DevOps shows how that experience carries over.

Section 3: Limits that are really in force

  • Record the memory limit and the process-count limit of every container. A blank means none.
  • If the host runs rootless containers, confirm the limits took effect. The Atlas notes that limit flags are ignored in rootless mode unless the machine uses cgroup version two with systemd, and that CPU limits need the controller delegated to the user.
  • Check recent history for kills with exit code 137, which equals 128 plus the kill signal 9. Note whether the cause was the container’s own ceiling or pressure on the whole host.
  • If CPU caps are set, read the throttling counters and compute throttled periods divided by total periods. Flag anything above 5 percent for a latency-sensitive service.

Section 4: Rootless mode, if you use it

Rootless operation is a strong safety measure and has trade-offs the Atlas lists. Verify that you have planned for them rather than discovering them during an outage.

  • Ports below 1024 need a system setting changed, or a different port behind a proxy.
  • Overlay networks are not supported, so multi-host designs need another approach.
  • User-mode networking is slower; measure it for any throughput-sensitive service.

Section 5: Stopping and logging

  • Read each start command. Shell form means the stop signal probably never reaches your program; exec form fixes that.
  • Time a stop. If every stop runs out the clock, with Compose’s default wait of ten seconds before the forced kill, your program is not reacting to the signal.
  • Check the logging driver settings. Out of the box the json-file driver never rotates, so look for a maximum size and a file count.
  • Check free disk space on the host and the size of the log directory.

Section 6: Names, images and ownership

  • List the image reference of every running container. Mark each one that uses a floating tag.
  • For each service, name the person or team who owns rebuilding it.
  • Record the age of each base image, and note which ones nothing is updating.
  • Record which registries the host pulls from and whether anonymous pulls are used. The Atlas quotes a limit of 100 anonymous pulls per six hours on Docker Hub, which busy hosts can exhaust.

Section 7: Evidence to keep

An audit is only as useful as its paper trail, so decide up front how you will store what you find. Two small habits make a large difference to the quality of the result, and both take minutes.

  • Save the output of each inspection command next to the notes. Screenshots go stale; text files can be compared later.
  • Record the date, the host name and the engine version, so a later audit can tell whether a change came from your fixes or from an upgrade.
  • Write the single most worrying finding at the top of the file. Readers rarely scroll.

A tidy audit file also guards you against a common pitfall: fixes that are made but never recorded, and so get quietly reverted at the next rebuild.

A short example of one finding, end to end

Imagine the port listing shows a database container published as 5432 to 5432 with no host address. Written up properly, the finding names the service, the exposed address, the way you confirmed it from a second machine and the fix: bind to loopback, route legitimate access through a proxy, and re-scan. The retest line is the important part, because it turns a claim into a demonstrated result.

Scoring and what to do next

Tally the failures under each heading. Whichever section holds the most gets your first week. Do not try to fix it all at once, because moving ports and switching users can interrupt live services. Make each change in a test copy first, then roll it out one service at a time and re-run the relevant check afterwards.

For your own records, store the audit as a dated file next to the Compose or build files. When you repeat it in three months, you will see which items improved, and the comparison makes a persuasive case when you ask for time to keep going. If you are building a career around this kind of work, the roadmap covering skills, projects and certifications is a good place to place this audit as a portfolio piece, and the overview of why certification helps can help you decide how formal to make your study.

Finish by noting what the audit does not cover. Image signing, vulnerability scanning depth, admission policies on a cluster and the host’s own kernel settings belong to neighbouring areas. Naming those gaps in the report is part of doing an honest audit.