harden: block on-demand autoload of tc classifiers, actions and exotic qdiscs

README.md

@@ -240,7 +240,15 @@ Three things worth knowing before pointing this at code you don't trust:
   is the security boundary here, not the guest's own user separation. Disabling
   them (`security.allowUserNamespaces = false`) would close it but breaks the Nix
   sandbox inside the guest, and `security.lockKernelModules` conflicts with
-  `ENABLE_CRI`, so neither is on by default.
+  `ENABLE_CRI`, so neither is on by default. What the guest does do is refuse the
+  on-demand autoload of the `net/sched` modules listed in
+  `blockedTcModules` in `modules/base.nix`, which removes the most travelled
+  route into `net/sched` without the `lockKernelModules` conflict. This is
+  guest-internal defence in depth: `modprobe.d` constrains modprobe-mediated
+  loads, not a direct `finit_module` from something already privileged in the
+  guest. It narrows the path to guest root; it does not close it, and the VM
+  remains the security boundary. If you add the CNI `bandwidth` plugin to the
+  chain, drop `act_mirred` and `cls_u32` from the list.
 - **On a single-user Nix install, the read-only store share is the only thing
   protecting the host store.** The `ro-store` share is exported `readOnly`, and on
   NixOS or a multi-user install the host store is additionally not writable by the

modules/base.nix

@@ -1,6 +1,84 @@
 { pkgs, lib, config, ... }:
 let
   cfg = config.claude-vm.agent;
+
+  # tc classifiers and actions, blocked from on-demand autoload below.
+  #
+  # Only modules that still exist upstream are listed. cls_tcindex and cls_rsvp
+  # were retired from the kernel (6.3 and later), so entries for them would be
+  # inert. cls_route survived that cull — it is still built and still modular
+  # (verified against /run/booted-system/kernel-modules/…/net/sched on 6.18.45),
+  # so it belongs on the list.
+  #
+  # em_* (ematch) and act_meta_* are deliberately absent, but not because
+  # entries would be inert: both are alias-loaded (ematch-kind-N, ife-meta-*)
+  # and modprobe applies `install` after alias resolution, so listing them would
+  # take effect. They are omitted because they are only reachable through
+  # cls_basic / cls_flow and act_ife, which are blocked here already. If any of
+  # those ever comes off the list, add the matching em_* / act_meta_*
+  # entries.
+  blockedTcFilters = [
+    "cls_u32"
+    "cls_fw"
+    "cls_basic"
+    "cls_flow"
+    "cls_cgroup"
+    "cls_flower"
+    "cls_matchall"
+    "cls_bpf"
+    "cls_route"
+    "act_pedit"
+    "act_mirred"
+    "act_police"
+    "act_gact"
+    "act_bpf"
+    "act_connmark"
+    "act_csum"
+    "act_ct"
+    "act_ctinfo"
+    "act_ife"
+    "act_mpls"
+    "act_nat"
+    "act_sample"
+    "act_simple"
+    "act_skbedit"
+    "act_skbmod"
+    "act_tunnel_key"
+    "act_vlan"
+    "act_gate"
+  ];
+
+  # Queueing disciplines, same treatment. Blocking the filters while leaving
+  # every sch_* faultable would leave the larger half of net/sched open, and the
+  # qdisc side has been at least as productive for local privilege escalation
+  # as the action side — sch_qfq alone accounts for CVE-2023-4921 and
+  # CVE-2023-31436.
+  #
+  # Only the exotic ones are listed. These are deliberately left loadable
+  # because they are actually used:
+  #
+  #   sch_tbf               — the CNI `bandwidth` plugin's ingressRate path
+  #   sch_ingress           — also provides clsact, which container networking
+  #                           and any tc-BPF attachment needs; the stronger
+  #                           reason it must stay loadable
+  #   sch_fq_codel          — net.core.default_qdisc, loaded on every boot and
+  #                           the only sched module live on a default guest
+  #
+  # All of them were present and modular on 6.18.45; none is reachable in
+  # normal use of this VM.
+  blockedTcQdiscs = [
+    "sch_qfq"
+    "sch_choke"
+    "sch_teql"
+    "sch_dualpi2"
+    "sch_cbs"
+    "sch_taprio"
+    "sch_etf"
+    "sch_plug"
+    "sch_skbprio"
+  ];
+
+  blockedTcModules = blockedTcFilters ++ blockedTcQdiscs;
 in
 {
   options.claude-vm.agent = {
@@ -92,6 +170,44 @@ in
 
     boot.kernelParams = [ "console=hvc0" ];
 
+    # Block on-demand autoload of tc classifiers, actions and the exotic
+    # queueing disciplines.
+    #
+    # Unprivileged user namespaces stay enabled (see the hardening notes in the
+    # README), so an unprivileged guest user holds namespaced CAP_NET_ADMIN and
+    # can reach net/sched. Loading is what makes that reach useful: the kernel
+    # pulls these in on first use via request_module(), so a guest that never
+    # legitimately touches tc can still fault in a classifier or action and
+    # attack it. Refusing the load closes the route as a category rather than
+    # one CVE at a time.
+    #
+    # `install <mod> false` rather than boot.blacklistedKernelModules: the
+    # latter emits nothing but `blacklist <name>` lines, which suppress
+    # alias-based loading but not a request by real name. cls_api.c and
+    # act_api.c ask through request_module("cls_%s") / ("act_%s") with the
+    # literal name, which a blacklist line does not stop. Please don't
+    # "simplify" this back.
+    #
+    # The command must be an absolute store path, not /bin/false: modprobe runs
+    # it through /bin/sh -c, and the guest's /bin holds exactly one entry (sh).
+    # /bin/false would exit 127 "command not found" -- the load is still refused,
+    # but by accident rather than by design, and it logs a misleading error on
+    # every attempt.
+    #
+    # Nothing in the default CNI chain (bridge + portmap + firewall) uses tc,
+    # so this is inert for ENABLE_CRI as shipped. It is compatible with
+    # container runtimes in a way security.lockKernelModules is not, since that
+    # sets kernel.modules_disabled=1 and blocks the on-demand loads CNI does
+    # need.
+    #
+    # If you add the `bandwidth` plugin to the chain, drop act_mirred and
+    # cls_u32 from the list: its egressRate path attaches a u32 filter carrying
+    # a mirred TCA_EGRESS_REDIR action to redirect into an ifb device (see
+    # CreateEgressQdisc in plugins/meta/bandwidth/ifb_creator.go upstream). Its
+    # ingressRate path only needs sch_tbf and is unaffected.
+    boot.extraModprobeConfig =
+      lib.concatMapStrings (m: "install ${m} ${pkgs.coreutils}/bin/false\n") blockedTcModules;
+
     services.getty.autologinUser = "agent";
     systemd.services."getty@tty1".enable = false;
 

scripts/changelog.sh

@@ -18,7 +18,7 @@ if [ -z "$range" ]; then
   range="${prev:+$prev..}HEAD"
 fi
 
-breaking=() feats=() fixes=() docs=() other=()
+breaking=() security=() feats=() fixes=() docs=() other=()
 
 # Records are \x1e-separated and fields \x1f-separated: commit bodies are
 # multi-line, so a plain line-oriented read would split them across records.
@@ -42,6 +42,8 @@ while IFS= read -r -d $'\x1e' record; do
     breaking+=("$entry")
   else
     case "$type" in
+      harden|sec|security)
+                  security+=("$entry") ;;
       feat)       feats+=("$entry") ;;
       fix)        fixes+=("$entry") ;;
       docs)       docs+=("$entry")  ;;
@@ -59,6 +61,7 @@ section() {
 }
 
 section "Breaking changes" ${breaking+"${breaking[@]}"}
+section "Security"         ${security+"${security[@]}"}
 section "Features"         ${feats+"${feats[@]}"}
 section "Fixes"            ${fixes+"${fixes[@]}"}
 section "Documentation"    ${docs+"${docs[@]}"}