How to harden a Redpanda container (and is redpanda:v24.2.11 safe for untrusted workloads?)¶
A streaming broker sits between every producer and consumer you run, which is exactly why its
container should be one of the tightest. A stock docker run redpandadata/redpanda:v24.2.11 is
not that. Graded on IronClaw's seven-dimension containment scale, the default configuration scores
63 of 100, grade C (partial). Higher is safer. A couple of runtime flags take the same image
to 89 of 100, grade B, one point off an A, and the one dimension it cannot reach is the one a
broker needs by definition (clients must connect to it). Here are the exact gaps and fixes from
the scan data.
Graded from a read-only inspect of a running container started from
redpandadata/redpanda:v24.2.11with plaindocker rundefaults, its entrypoint overridden withsleeppurely to keep it alive. The scan itself executes nothing inside the container. It is the same data behind its isolation scorecard. How scoring works →
Where the default configuration leaks¶
ironctl scan grades seven independent containment boundaries. On a default
docker run redpandadata/redpanda:v24.2.11, three fail or warn:
| Dimension | Verdict | Score | What the scan found |
|---|---|---|---|
| Non-root user (uid != 0) | ✅ PASS | 15/15 | runs as redpanda (uid != 0) |
| Dropped capabilities | ❌ FAIL | 4/20 | default capability set retained (CAP_NET_RAW, CAP_MKNOD, and more) |
| Seccomp profile | ✅ PASS | 15/15 | seccomp profile active |
| Network isolation / egress | ⚠️ WARN | 4/15 | network=bridge: outbound egress is possible |
| Read-only root filesystem | ❌ FAIL | 0/10 | root filesystem is writable |
| No docker.sock exposure | ✅ PASS | 15/15 | no control socket mounted |
| No shared host namespaces | ✅ PASS | 10/10 | no host PID/IPC/network sharing |
Redpanda already runs as a non-root uid, which is why it clears the default 48/100 that root
images sit at. The sharpest remaining edges are the retained capabilities and open egress:
a Kafka-protocol or admin-API CVE that lands code execution lands it with CAP_NET_RAW and
friends, on a process every service in your stack already trusts and connects to, and it can reach
the network to exfiltrate every message on the log.
Harden it: the exact --fix remediation¶
ironctl scan my-redpanda --fix prints one remediation per failed dimension, then one hardened
run. For redpandadata/redpanda:v24.2.11:
--cap-drop=ALL(Dropped capabilities, +16): drop every Linux capability; add back only what the workload provably needs. Redpanda needs none of the defaults.- Scoped network (Network isolation):
--network=nonescores the full 15 but is wrong for a broker, producers and consumers must be able to connect to it. Any named or bridge network scores 4 of 15 (a WARN, not a fail): a connection path exists. This is the one dimension a broker cannot max out. Contain it anyway: attach a user-defined network scoped to just its producers and consumers, with no default route out, so a compromised broker cannot call arbitrary internet addresses. --user 65532:65532(Non-root user, already passing at +0): non-root already passes as the redpanda uid. Pinning an explicit fixed uid is the auditable choice and keeps volume ownership unambiguous; it does not change the score.--read-only --tmpfs /tmp(Read-only rootfs, +10): make the root filesystem read-only and mount/var/lib/redpanda/dataas an explicit writable volume. Removes the persistence surface.
Before and after¶
# Before: 63/100, grade C
docker run -d --name redpanda \
redpandadata/redpanda:v24.2.11 \
redpanda start --mode dev-container
# After: 89/100, grade B (scoped private network for producers and consumers)
docker run -d --name redpanda-hardened \
--user 65532:65532 \
--cap-drop=ALL \
--security-opt=no-new-privileges \
--read-only --tmpfs /tmp \
-v redpanda-data:/var/lib/redpanda/data \
--network=redpanda-internal \
redpandadata/redpanda:v24.2.11 \
redpanda start --mode dev-container
Rescan: ironctl scan redpanda-hardened reports 89/100 grade B. A 26-point swing with no
custom image build, just the right flags. The only dimension still short of full marks is the
network (4 of 15), because a broker exists to be connected to; network=none would score the
last points but leave nothing able to reach the log. That is the honest ceiling for a broker, and
it is a long way from the default C.
Verify it on your own broker¶
# install (Homebrew)
brew install ironsecco/ironclaw/ironclaw
# grade your running container, then print the fixes
ironctl scan my-redpanda
ironctl scan my-redpanda --fix
ironctl scan also reads a docker-compose.yml service or a Kubernetes manifest, so you can
grade the Redpanda in your stack, not just a bare docker run.
Keep going¶
- redpanda:v24.2.11 isolation scorecard →: the full dimension breakdown.
- Message queues, ranked by isolation →: how Redpanda compares to Kafka, RabbitMQ, NATS, and the rest.
- How to harden a Kafka container →: the same honest-ceiling walkthrough for the broker Redpanda speaks the protocol of.
- Scan any container in 10 seconds →: the full
ironctl scanreference. - Run untrusted code in a real sandbox →: IronClaw wraps every AI-agent session in a gVisor/Kata boundary with
network=noneby default.