Re: [PATCH v2 1/4] userns: introduce capability bounding set for user namespaces
From: Andy Lutomirski
Date: Fri Nov 20 2026 - 12:10:11 EST
> On Nov 19, 2026, at 2:52 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> On Thu, Nov 19, 2026 at 2:43 AM Hallgrimur H. Gunnarsson
> <halg@xxxxxxxxxx> wrote:
>>
>> When a process creates a new user namespace, it receives a full set of
>> capabilities (CAP_FULL_SET) scoped to that namespace. While namespaced
>> in scope, several of these capabilities (such as CAP_NET_ADMIN and
>> CAP_NET_RAW) expose broad kernel attack surface that many sandboxes and
>> containers never need. However, there is currently no configurable way
>> to withhold such capabilities from user namespaces while still granting
>> them to workloads that require them.
>
>
> One thought (from the rootns thread): the other decent sized attack
> surface here is CAP_MOUNT_ADMIN. Which doesn't exist as a capability
> bit.
>
> A sort of horrible solution (but maybe not so awful) would be to
> declare that the bits in cap_userns aren't actually capability bits,
> and that you need the bit in cap_userns *and* the bit in cap_effective
> to do sensitive things, and that mounting needs cap_userns to have
> NSCAP_MOUNT_ADMIN *and* cap_effective to have CAP_SYS_ADMIN.
>
> I want to sleep on this. It might actually solve several problems,
> e.g. controlling creation of ipv4 and ipv6 loopback or tunnel devices
> or whatever.
I slept on it. Here’s a counterproposal. I’m not saying I think this
is better than this patch set, but I think it’s worth weighting the
options.
Define a whole new set of privilege bits. As a
poorly-painted-bikeshed option, I’ll call them eprivs (for exposure
privileges - it’s a crappy name but at least it’s pronounceable).
An epriv is not permission to do something *to* an object; it’s
permission to do something that exposes a meaningful attack surface
within the kernel. So there’s no ns_has_epriv because there is no
object. These are kind of the opposite of object capabilities.
Additionally, an epriv never grants the ability to do something that
one couldn’t do in an epriv-less world.
For now, all tasks have all eprivs, even not-yet-defined ones, unless
user code or config explicitly changes this. (Maybe some day secure
boot might remove certain eprivs.)
There should never be a check like ns_capable || has_epriv. It’s also
&&, or perhaps caps are not checked at all.
Example eprivs include:
EPRIV_FSOPEN_VIRT: may fsconfig, fsopen, etc a virtual fs like fuse or
tmpfs or overlayfs.
EPRIV_CLONE_TREE: may clone a mount tree
EPRIV_MOVE_MOUNT: may move mounts around (but only if you also have
CAP_SYS_ADMIN on the correct namespace, of course)
EPRIV_FIREWALL_TC: may configure any firewalling configuration or tc
configuration (this includes iptables, nftables, and ebtables)
EPRIV_NET_RAW: may use raw/packet sockets (like CAP_NET_RAW)
EPRIV_TUNTAP: may create and attach to tun/tap interfaces
EPRIV_IP_LOOPBACK: may set up IPv4 and IPv6 loopback using whatever
new facility we add
EPRIV_NETLINK: may use netlink at all
EPRIV_NET_ADMIN: may use traditional CAP_NET_ADMIN abilities (all
CAP_NET_ADMIN checks will also check EPRIV_NET_ADMIN)
EPRIV_OPEN_FILE_HANDLE: May use open_by_handle_at, etc.
EPRIV_UNSHARE_USERNS: may unshare userns
EPRIV_UNSHARE_NONUSERNS: may unshare any other namespace
EPRIV_MOUNT_CGROUPv1, EPRIV_MOUNT_CGROUPv2: may create a new mount of
these subsystems
EPRIV_PTRACE: there are lots of different variations that might make sense.
EPRIV_MOUNT_PROC_PIDONLY: may mount proc-pidonly
EPRIV_MOUNT_PROC_NOTPIDONLY: may mount proc not-pidonly
EPRIV_MOUNT_SYSFS: may mount sysfs
Example eprivs that are questionable:
EPRIV_DMESG: the kernel log is a specific object, and I think it's
better to control access using something kernel-log-specific.
There is no operation that increases eprivs -- even without
no_new_privs set, a task will never gain an epriv, nor will a child
ever have an epriv that its creator did not have at the time the child
was created.
The evolution rules could be as simple as, on fork/clone, the child
inheriting exactly the parent's eprivs.
Or we could bind eprivs to a userns instead of a task, but that could
be more limiting and awkward than we want.
What do you all think?
FWIW, I had an LLM make an inventory (surely not exhaustive) of
recent-ish CVEs and how they relate to this list. (The below is the
only part of this email that is LLM-generated, although I added
EPRIV_SYSFS because I didn't realize that nonroot userns could mount
it until I asked an LLM, sigh.) To me, the list and its breakdown
suggests that there is quite a lot of room for meaningful attack
surface reductions while still leaving large parts of userns
available, and that my epriv proposal captures much of this but the
cap_userns proposal may be too coarse-grained.
(I also asked about eBPF, and my LLM came up with a lot of CVEs, but
there's an existing BPF token mechanism and I therefore did not
propose EPRIV_LOAD_EBPF.)
--- begin LLM output ---
**EPRIV_FSOPEN_VIRT:** CVE-2022-0185*, CVE-2023-0386*, CVE-2023-2640*,
CVE-2023-32629*, CVE-2025-38499*
**EPRIV_CLONE_TREE:** CVE-2025-38499* (only if the check also covers
the relevant internal mount-cloning path; explicit `open_tree()`
restrictions alone are insufficient)
**EPRIV_MOVE_MOUNT:** CVE-2025-37988
**EPRIV_FIREWALL_TC:** CVE-2022-25636, CVE-2022-1015, CVE-2022-1016,
CVE-2022-32250, CVE-2022-34918, CVE-2022-48911*, CVE-2022-2588,
CVE-2023-32233, CVE-2023-35001, CVE-2023-4004, CVE-2023-4015,
CVE-2023-4244, CVE-2023-5197, CVE-2023-6817, CVE-2023-3609,
CVE-2023-4623, CVE-2023-4921, CVE-2024-1086, CVE-2024-0607,
CVE-2024-27397, CVE-2024-26925, CVE-2025-38001, CVE-2026-23111
**EPRIV_NET_RAW:** None identified.
**EPRIV_TUNTAP:** None identified as requiring TUN/TAP. Some
packet-triggered exploits may optionally use it, but that does not
establish it as a necessary blocker.
**EPRIV_IP_LOOPBACK:** None identified.
**EPRIV_NETLINK:** Also those blocked by `EPRIV_FIREWALL_TC`, insofar
as the documented attack paths use netlink. No additional CVEs
identified.
**EPRIV_NET_ADMIN:** Also those blocked by `EPRIV_FIREWALL_TC`,
insofar as the vulnerable operations require `CAP_NET_ADMIN`. No
additional CVEs identified.
**EPRIV_OPEN_FILE_HANDLE:** CVE-2025-38306*
**EPRIV_UNSHARE_USERNS:** Also those blocked by `EPRIV_FSOPEN_VIRT`,
`EPRIV_MOVE_MOUNT`, `EPRIV_FIREWALL_TC`, `EPRIV_OPEN_FILE_HANDLE`,
`EPRIV_MOUNT_CGROUPv1`, and `EPRIV_MOUNT_SYSFS`, plus CVE-2025-38058*.
More generally, blocks the documented unprivileged user-namespace
setup for the entire inventory when no usable user namespace already
exists.
**EPRIV_UNSHARE_NONUSERNS:** Also the namespace-dependent attack paths
blocked by `EPRIV_FIREWALL_TC`, `EPRIV_MOVE_MOUNT`,
`EPRIV_MOUNT_CGROUPv1`, `EPRIV_MOUNT_SYSFS`, and many of those blocked
by `EPRIV_FSOPEN_VIRT`. Coverage depends on whether the necessary
non-user namespaces already exist.
**EPRIV_MOUNT_CGROUPv1:** CVE-2022-0492*
**EPRIV_MOUNT_CGROUPv2:** None identified.
**EPRIV_PTRACE:** None identified.
**EPRIV_LOAD_EBPF:** None identified.
**EPRIV_MOUNT_PROC_PIDONLY:** None identified.
**EPRIV_MOUNT_PROC_NOTPIDONLY:** None identified.
**EPRIV_MOUNT_SYSFS:** CVE-2025-22089*
CVEs not cleanly blocked by an operation-specific bit
Excluding the two EPRIV_UNSHARE_* gates:
- CVE-2025-38058: No explicit unmount privilege. The vulnerable race
involves umount(2) and path lookup.
- CVE-2022-0492: The cgroup-v1 mount bit covers the documented
new-mount attack path, but not necessarily configuration of
release_agent on an existing writable hierarchy.
- CVE-2025-38499: The correct gate depends on which filesystem-context
or internal cloning operation reaches clone_private_mnt(). I would not
count EPRIV_CLONE_TREE as a definitive blocker without checking the
exact call paths.