4.8 KiB
status, opened, located-in, fixed-by, amended-design
| status | opened | located-in | fixed-by | amended-design | |
|---|---|---|---|---|---|
| resolved | 2026-09-14 |
|
mesh-controller c8d8211 (forward chain drops by default, never closes ssh), 25e42b3; proven by one-node-mesh.test.ts (a machine filters exactly what its modules declared) |
047 — The firewall does not cover published container ports
Symptom
A machine running the mesh's firewall drops what it should drop and leaves every container port published on all addresses open to anything that can reach the machine.
Rehearsed in the lab on three machines, one probing another, with a single change between the two measurements:
| ordinary port on the host | container port published on 0.0.0.0 |
|
|---|---|---|
| before the mesh's rules | reachable | reachable |
| after the mesh's rules | refused | still reachable |
Same prober, same ports, same second. The first column is the firewall working. The second is the question.
Why this is the way it is
It is deliberate, and the reasoning is written where the rules are generated: the mesh's table hooks
input and output and no forward chain, because what a machine forwards is the container
runtime's business — the runtime writes rules for the bridges it creates, and a second table
dropping in forward would drop container traffic the runtime explicitly allowed, stopping every
container on the node including the control plane.
That reasoning is sound. The conclusion drawn from it is not the only one available, because traffic
to a published port does not traverse input at all: it is redirected and then forwarded. So a
default-drop input chain says nothing whatsoever about it, and the firewall is silent on the ports
most likely to matter.
Why it matters, measured rather than argued
The anchor this mesh will be installed onto publishes 38 distinct container ports on all addresses — a document store, three cache sentinels, mail transfer and retrieval, the forge's own SSH, and thirty more.
Today those are filtered, by the system being replaced, which does it with rules that hook exactly the path this one leaves alone. So the cutover from the old firewall to this one opens all thirty-eight at once, silently, on a machine with a public address, while every check reports a firewall that is up and dropping by default.
That is the worst shape a security fault can take: the mechanism is present, it reports success, and the thing it is believed to be doing is not the thing it does.
The other half, which runs the opposite way
The section above is about ports the firewall cannot close. There is a matching fault about ports it does close and should not, and together they are worse than either alone.
The rules are computed from what modules declare they listen on. During a migration a mesh knows about almost nothing — the services are still running under the system being replaced — so it opens almost nothing. Anything listening on the host rather than in a container is then dropped.
Including ssh. No module in the catalogue declares an ssh port, and the generator has no built-in allowance for one. So the ruleset loaded on a machine that has not been told otherwise accepts established connections, loopback and ping, and refuses every new ssh connection.
The session doing the loading survives, because established connections are accepted. It survives until it drops. On a machine reached over the network that is the difference between a mistake and a journey to a rescue console.
So the two halves are:
| what it does | consequence | |
|---|---|---|
| container ports, published | cannot see them, does not filter | stay open, protection silently lost |
| host ports, undeclared | sees them, drops them | ssh among them |
The mesh gets no say over the ports most worth protecting, and full say over the one that must never be closed by accident.
How it would be checked
Two probes from another machine, against a port on the host and a published container port, before and after the rules are loaded. Both must change together. The lab can do this — it is how the table above was produced — and nothing does it today.
Open questions
- Can the mesh filter forwarded traffic without owning the whole
forwardchain? A table that accepts by default and drops only what the mesh knows about would not stop the runtime's traffic, and hookingforwardat all is the thing the current comment rejects. - Should a module's declared reach apply to its published ports? A module that says it listens to the mesh only, and is then reachable from everywhere because its port is published, is not doing what its manifest says.
- Is the answer instead that the mesh should not publish on all addresses? A port bound to the private network needs no filtering, and most of these want exactly that.