Docker security non-events
Estimated reading time: 3 minutes
This page lists security vulnerabilities which Docker mitigated, such that
processes run in Docker containers were never vulnerable to the bug—even before
it was fixed. This assumes containers are run without adding extra capabilities
or not run as
The list below is not even remotely complete. Rather, it is a sample of the few bugs we’ve actually noticed to have attracted security review and publicly disclosed vulnerabilities. In all likelihood, the bugs that haven’t been reported far outnumber those that have. Luckily, since Docker’s approach to secure by default through apparmor, seccomp, and dropping capabilities, it likely mitigates unknown bugs just as well as it does known ones.
The introduction of unprivileged user namespaces lead to a huge increase in the
attack surface available to unprivileged users by giving such users legitimate
access to previously root-only system calls like
mount(). All of these CVEs are examples of security vulnerabilities due to introduction of user namespaces. Docker can use user namespaces to set up containers, but then disallows the process inside the container from creating its own nested namespaces through the default seccomp profile, rendering these vulnerabilities unexploitable.
These are bugs that require the presence of a setuid binary. Docker disables
setuid binaries inside containers via the
NO_NEW_PRIVSprocess flag and other mechanisms.
A bug in
ptrace()could allow privilege escalation. Docker disables
ptrace()inside the container using apparmor, seccomp and by dropping
CAP_PTRACE. Three times the layers of protection there!
A series of crafted
keyctl()calls could cause kernel DoS / memory corruption. Docker disables
keyctl()inside containers using seccomp.
4036: These are
bugs in common virtualization drivers which could allow a guest OS user to
execute code on the host OS. Exploiting them requires access to virtualization
devices in the guest. Docker hides direct access to these devices when run
--privileged. Interestingly, these seem to be cases where containers are “more secure” than a VM, going against common wisdom that VMs are “more secure” than containers.
Use-after-free caused by crafted
keyctl()calls could lead to privilege escalation. Docker disables
keyctl()inside containers using the default seccomp profile.
A bug in eBPF -- the special in-kernel DSL used to express things like seccomp
filters -- allowed arbitrary reads of kernel memory. The
bpf()system call is blocked inside Docker containers using (ironically) seccomp.
A bug in setsockopt with
ARPT_SO_SET_REPLACEcausing memory corruption / local privilege escalation. These arguments are blocked by
CAP_NET_ADMIN, which Docker does not allow by default.
Bugs not mitigated:
5157: Bugs in
the kernel’s non-maskable interrupt handling allowed privilege escalation.
Can be exploited in Docker containers because the
modify_ldt()system call is not currently blocked using seccomp.
A race condition was found in the way the Linux kernel’s memory subsystem
handled the copy-on-write (COW) breakage of private read-only memory mappings,
which allowed unprivileged local users to gain write access to read-only memory.
Also known as “dirty COW.”
Partial mitigations: on some operating systems this vulnerability is mitigated
by the combination of seccomp filtering of
ptraceand the fact that