How to harden a Neo4j container (and is neo4j:5 safe for your graph?)¶
Neo4j holds a connected picture of your data: identity graphs, fraud rings, recommendation edges, and
the relationships an attacker would love to walk. A stock docker run neo4j:5 keeps that graph behind
a boundary weaker than the data deserves. Graded on IronClaw's seven-dimension containment scale, the
default configuration scores 48 of 100, grade D (porous). Higher is safer. Unlike a broker or a
proxy, a graph database that only its co-located application talks to can close every dimension,
including the network. A few runtime flags take the same image to a full 100 of 100, grade A. Here
are the exact gaps and fixes from the scan data.
Every number here comes from a read-only
docker inspectofneo4j:5, the same data behind its isolation scorecard. No workload is executed. How scoring works →
Where the default configuration leaks¶
ironctl scan grades seven independent containment boundaries. On a default docker run neo4j:5,
three fail and one warns:
| Dimension | Verdict | Score | What the scan found |
|---|---|---|---|
| Non-root user (uid != 0) | ❌ FAIL | 0/15 | runs as root (uid 0); a container escape starts with host-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 |
The two that should worry you most are root and egress. A Neo4j process that escapes as root escapes as root on the host, next to the very store files it was holding. And a database that can reach arbitrary destinations is one that can quietly ship your entire graph out the moment a Cypher parsing or plugin CVE lands code execution. The default capability set and writable rootfs widen and entrench that foothold.
Harden it: the exact --fix remediation¶
ironctl scan my-neo4j --fix prints one remediation per failed dimension, then one hardened run. For
neo4j:5:
--user 7474:7474(Non-root user, +15): pin the non-rootneo4juid so an escape does not begin as host uid 0. Point the data and logs directories at a volume this uid owns.--cap-drop=ALL(Dropped capabilities, +16): drop every Linux capability; Neo4j needs none of the default set to serve Bolt and HTTP on high ports.--read-only --tmpfs /tmp(Read-only rootfs, +10): make the root filesystem read-only and mount/dataand/logsas explicit writable volumes. Removes the persistence surface.--network=none(Network isolation, +11 to the full 15): this is the dimension a co-located store can actually max out. If the only client is an application on the same host or pod reaching Neo4j over the loopback of a shared network namespace, cut the NIC entirely. Nothing external can connect, and the database cannot phone home.
When network=none is not honest¶
If remote clients or Neo4j browser users connect over the network, or you run a causal cluster with
peer connections between core members, you cannot use --network=none; the store has to accept those
connections. In that case put it on a user-defined network scoped to just its clients and cluster
peers, with no default route out. That holds the network dimension at a WARN (4 of 15) and the honest
ceiling becomes 89 of 100, grade B, the same as a broker. Use --network=none only for the
single-application, co-located case.
Before and after¶
# Before: 48/100, grade D
docker run -d --name neo4j neo4j:5
# After: 100/100, grade A (co-located store, no network needed)
docker run -d --name neo4j-hardened \
--user 7474:7474 \
--cap-drop=ALL \
--security-opt=no-new-privileges \
--read-only --tmpfs /tmp \
-v neo4j-data:/data -v neo4j-logs:/logs \
--network=none \
neo4j:5
Rescan: ironctl scan neo4j-hardened reports 100/100 grade A. A 52-point swing with no custom
image build, just the right flags. Every dimension is closed because a co-located graph database does
not need to talk to anything but the app on the other side of its loopback. That is the top grade,
reserved for datastores whose clients live next to them.
Verify it on your own Neo4j¶
# install (Homebrew)
brew install ironsecco/ironclaw/ironclaw
# grade your running container, then print the fixes
ironctl scan my-neo4j
ironctl scan my-neo4j --fix
ironctl scan also reads a docker-compose.yml service or a Kubernetes manifest, so you can grade
the Neo4j in your stack, not just a bare docker run.
Keep going¶
- All hardening guides →: every harden-a-container walkthrough, with grade deltas.
- neo4j:5 isolation scorecard →: the full dimension breakdown.
- How to harden a MongoDB container →: another document-shaped store that reaches grade A when co-located.
- 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.