The Golden Ticket: Docker and High Security Microservices by Aaron Grattafiori

The Golden Ticket: Docker
and High Security
Microservices
Aaron Grattafiori
Technical Director

whoami(1)
Longtime Hacker,
Pentester, and Linux geek
Wrote a few things about
Linux Containers…

a disclaimer…
Microservices are not for everyone
Microservices can be hard
Microservices often distribute complexity

You’ve seen Microservices before

“Worker bees can leave.
Even drones can fly away.
The Queen is their slave.”

int main() {
printf(“foon”);
return 0;
}
void DoSomething() {
printf (“foon”);
}
int main() {
DoSomething();
return 0;
}

A layered defense…
Shrink attack surfaces and present
harden those which remain
The Principle of Defense in Depth

Monolithic applications make this difficult
and Microservices make this easy
Key: not having a single point of security failure
The Principle of Defense in Depth

“Absolute power corrupts absolutely”
“Avoid running as root” is about
shrinking trust boundaries.
Monoliths make this difficult and
Microservices make this easier
The Principle of Least Privilege
#

“Sane defaults, Isolate by trust”
The Principle of Least Surprise

“Access provided on a need to know basis”
Military security, Mafia management,
database connections, defense in depth, …
The Principle of Least Access

This will be become a theme, a war against excess and
complexity…
Microservices have complexity “at scale”
Microservices still reduce overall complexity
“Least” is Common to all…

1)Establish trust boundaries
2)Identify, minimize and harden attack surfaces
3)Reduce scope and access
4)Layer protections / defenses
See “Stop Buying Security Prescriptions” by Justin Schuh of Google
.. Because “Least” is a Good Idea!

Upsides of Monolith AppSec
Building and Standing it up is a “known known”
Architecture is quite simple
Existing culture of dev/business/compliance/sales

Downsides of Monolith AppSec
Compromise of a single point often means compromise
of the entire application or network
Authentication requirements/creds are global in scope
Simple to build, but security is hard to tailor

Upsides of Microservices AppSec
UNIX model works well
Highly application/function specific containers permit
highly application specific security
Establishing a Trusted Computing Base (TCB)

Upsides of Microservices AppSec
Least privilege and app-specific security is much easier
Greatly reduced attack surfaces
Independent patching/updates/versions

Downsides of Microservices AppSec
A “Known unknown”
Requires good understanding of application
Legacy applications are not easily adapted

Downsides of Microservices AppSec
May require devops and other culture changes
Complexity breeds insecurity
“Requires” dumb pipes networking architecture

Exploring Real World Compromise
“Imagetragick” RCE
“Shellshock” RCE
General Command Injection
General LFI/LFR
Denial of Service

API Web Service
Front-end
Presentation
Back-end
Application
Database
LB

API Web Service
Front-end
Presentation
Back-end
Application
Database

Relationship Status: “Its Complicated”

Exploring Real World Solutions
API Web Service
Front-end Presentation
Back-end Application
Database

Enforcing Security Boundaries
API Web Service
Front-end Presentation
Database
LB

Breaking Up Again
Payments
AuthZ
Uploads
Cache
Legacy

Limit Compromises
Payments
AuthZ
Uploads
Cache
Legacy
Nope
Nope
Nope
Limited

Limit Compromises: OSI Edition

Layer 7 Authentication: Application

Layer 4/5 (7) Authentication: TLS

Thinking about it another way
What’s the attack surface of:
an SSL VPN using username+pw auth
vs.
an SSL VPN using TLS client certs
vs.
an IPSEC-only endpoint using key-based auth ?

Thinking about it another way
Layer 7 Authentication/Encryption? aka XML
Encryption or HTTP Header auth
Or Layer 4/5 Authentication/Encryption? aka TLS/mTLS
Or Layer 3 Authentication/Encryption? aka IPSEC

In general…
Layer 7 authentication should be required
Less than layer 7 authentication is highly desirable
Don’t forget Authentication != Authorization
Don’t forget about transport security

Containers Map to Microservices
Container Model Microservices Model
Root Capabilities Capability based security
Different network namespaces Overlay networks or SDN
One application per container One core function per application
… …

A Few Attacks of Container Systems
Cross Container Attacks on Host
Cross Container Attacks on Network
Exploiting Vulnerable Application
Exploiting Vulnerable Library
Exploiting Out of Date Docker Engine
Exploiting Host Kernel
Targeting Container Configuration
Poisoning Service Discovery
Targeting Docker API Configuration
Attacking Container Management

Pruning The Attack Tree
Application attack? Container itself,
capabilities, immutable files, mount flags,
MAC, read-only rootfs, defensive coding, etc.
Kernel/syscall exploit ?
Seccomp and kernel hardening
Seccomp weakness?
Manditory Access Control

Pruning The Attack Tree
Compromised kernel/host OS?
Network hardening, isolation on trust,
logging, least privilege, least access
Compromised employee credentials?
Least access, least privilege, key-based
authentication

Microservices in
Docker or runc

Why Docker / runc ?
Helps with standardization, testing, security
Strong security defaults (seccomp, caps) and options (user namespace)
App containers help follow and develop the Microservice model

Why use runC for MS?
Lightweight and minimal!
Open Container Initiative
Docker’s libcontainer

Why not runC ?
Lightweight and minimal/manual
Weak documentation/examples/debugging
Warning: Apparmor not a default build tag and Apparmor
support / User namespace support not enabled by default

Security starts with the base OS
Minimal Distro (CoreOS/RancherOS/etc)?
Plain old Ubuntu Server?
Clear Linux?
… Unikernel?

Security starts with the base OS
It’s important to understand how a distribution:
handles updates,
binary package compilation,
security defaults (MAC),
default kernel options,
sysctl settings, ….

Minimal: Kernel
Common sense configuration
make menuconfig: configuration, modules
grsecurity

Minimal Container?
… Why stop at shrinking just the host?
Don’t we want least access? Less updates?

Minimal: Container Images
docker-nginx mongo
FROM debian:jesse FROM debian:wheezy
redis node
FROM debian:jesse FROM buildpack-deps:jessie

So what exactly is FROM debian:jesse ?
FROM scratch
ADD rootfs.tar.xz /
CMD ["/bin/bash"]
= 137 MB

So what about FROM debian:wheezy ?
97 MB!
So what about Jenkins? ElasticSearch? Postgres?
FROM debian:jesse

Not Minimal Enough!
Even before an apt-get, there are still tons of libraries,
executables, perl, even language files our process doesn’t need!
That’s more patching, disk space, attack surface and post-
exploitation utilities we don’t need

General idea for Docker
$ cat Dockerfile
FROM scratch
ADD stuff
CMD [“/stuff”]
$ docker build

General idea for runC
$ mkdir rootfs
$ cp stuff rootfs/
$ runc spec
$ sudo runc run foo

Reality for runC (and Docker)
$ ldd /path/to/program
$ mkdir –p rootfs /usr /lib # … etc
$ cp /path/to/program rootfs/
$ cp /path/to/lib.so.0 rootfs/lib/
$ *cross fingers*

Minimal Container Images: Go + Docker
https://blog.codeship.com/building-minimal-
docker-containers-for-go-applications/http://blog.xebia.com/create-the-smallest-possible-docker-container/

Golang wiki server example
$ ls -l rootfs
rootfs/wiki
rootfs/usr/lib/libpthread.so.0
rootfs/usr/lib/libc.so.6
rootfs/lib64/ld-linux-x86-64.so.2
$ du –hs rootfs
7.3M

Minimal Container Images: Nginx + Docker
Roughly 32 libs + /usr/sbin/nginx + a few config
files and dirs…
= 15 MB!

“Make things as simple as
possible, but not simpler.”
— Albert Einstein

nanosleep(2)
Minimal (hardened) Host OS: ✔
Minimal (hardened) Root FS: ✔
Minimal (hardened) Kernel: ✔

Mandatory Access Control
MAC as a Linux Security Module (LSM)-
AppArmor, SELinux, etc. Grsecurity RBAC is also
nice
OSX uses TrustedBSD, Microsoft has MIC

Mandatory Access Control
Default Docker AppArmor policy is very good
Monolithic MAC policy must include a large number of
files, permission grants, complexity
Microservices again allow for specific security

Nested AppArmor
Profile /bin/bash {
/bin/cat Px,
profile /bin/cat {
/etc/hosts r,
}
}

Going back to Golang HTTPd PoC
Profile /wiki {
network inet stream,
/usr/lib/* mr,
/lib64/* mr,
/wiki mixr,
}

Custom AppArmor Profiles
Generating profiles with aa-genprof
Understanding/using application
is key, profiling is still required

Common mistakes/problems:
- Providing too much access
- Dealing with wildcards
- Path-based ACLs

Easier Mode for Docker/runC: Using Bane for profiles
by Jessie Frazelle
Hard Mode: FullSystemPolicy
Avoid deny lists

AppArmor Profile Gotchas
Profiles must be loaded by AppArmor first
Abstractions may be more verbose than you would like
Exercise your app is key, run unit/QA/regression tests
Profiling is difficult within a container itself

Seccomp Profiles
MAC is vulnerable to kernel attacks… kernel is huge attack surface
Seccomp default filter enabled in Docker Engine >= 1.10 !!
Seccomp is widely used in security, recently now with BPF

Putting the BPF into Seccomp BPF
General BPF pseudocode:
1) Check architecture
2) Get syscall number
N) ALLOW or DENY syscall
N+1) KILL, TRAP, TRACE, ALLOW process

Why Custom Profiles?
Default seccomp filter permits 304 syscalls :/
Minimal applications need minimal syscall set
We want to practice least privilege!

Methods for Generating Seccomp Profiles
1) strace/ltrace
2) Kernel help (sysdig or systemtap)
3) Auditd/auditctl
4) Seccomp itself w/ SECCOMP_RET_TRACE and
PTRACE_O_TRACESECCOMP

Seccomp Profiles using strace
strace via ptrace(2)
Weird timing bugs can occur, count will miss some calls
Tracing outside of Docker should be fine
Some programs just don’t want to be traced

strace (and also ltrace –S)
$ strace –f /bin/ls
execve("/bin/ls", ["ls"], [/* 21 vars */]) = 0
brk(0) = 0x8c31000
mmap2(NULL, 8192, PROT_READ, MAP_PRIVATE|MAP_ANONYMOUS, -1,
0) = 0xb78c7000
access("/etc/ld.so.preload", R_OK) = -1 ENOENT (No
such file or directory)
open("/etc/ld.so.cache", O_RDONLY) = 3
fstat64(3, {st_mode=S_IFREG|0644, st_size=65354, ...}) = 0
…

Seccomp Profiles using sysdig
Sysdig is cool! Really cool
Requires a kernel module :/
But it’s really cool!

sysdig
$ sysdig evt.type=open and proc.name=niginx
$ sysdig –p “%evt.type” proc.name=niginx
$ sysdig –p “%evt.type” | sort –u >
nginix.syscalls.txt

Seccomp Profiles using auditd
Auditd is a nice system for… auditing
Requires a daemon and complex-ish config
Very cool power built-in but not ideal for this!
(Could collect calls over time )

Seccomp Profiles using Seccomp
Basic C example by Kees Cook and Will Drewry:
(https://outflux.net/teach-seccomp/)
– Profiling via SECCOMP_RET_TRAP is slow going
Using SECCOMP_RET_TRACE is better!

But why?
ALLOW, TRAP, TRACE, KILL, ERRNO are limited,
all of them block the call except for TRACE
(and default ALLOW is not a good model)

- Create a given syscall whitelist
- Set the default action to SECCOMP_RET_TRACE
- use ptrace(2) with PTRACE_O_TRACESECCOMP
- Log appropriate info
- Repeat until no soft-failures!

Ptrace PoC
$ seccomp-trace –f filter.txt /path/to/program
...
called sycall # 157 : prctl
called sycall # 157 : prctl
SECCOMP WOULD BLOCK 56
called sycall # 56 : clone
called sycall # 14 : rt_sigprocmask
called sycall # 13 : rt_sigaction

General Seccomp Pitfalls
Profiles are fragile!
Profiles are architecture dependent- limiting
container portability

Seccomp in Docker
Default (large) whitelist or:
--security-opt seccomp=<profile>
"seccomp": {
"defaultAction": "SCMP_ACT_KILL",
"architectures": [ "SCMP_ARCH_X86" ],
"syscalls": [
{ "name": "getcwd",
"action": "SCMP_ACT_ALLOW"
}
]
}

Seccomp in runC
Part of the default build!
"seccomp": {
"defaultAction": "SCMP_ACT_KILL",
"architectures": [ "SCMP_ARCH_X86" ],
"syscalls": [
{ "name": "getcwd",
"action": "SCMP_ACT_ALLOW"
}
]
}

Seccomp notes
Remember that seccomp filters are not sandboxes
Libseccomp is nice and has Go bindings
see also: oz sandbox train mode by Subgraph OS
Be careful implementing it from scratch, blacklists
/denylists are risky! (don’t allow ptrace!)

Seccomp Pro Tips for Docker/runC
https://github.com/konstruktoid/Docker/blob/master/
Scripts/genSeccomp.sh can save time or JSON pain !
You can always confirm Seccomp by looking in:
grep Seccomp /proc/<PID>/status

nanosleep(2)
Minimal (hardened) Host OS: ✔
Minimal (hardened) Root FS: ✔
Minimal (hardened) Kernel: ✔
Minimal AppArmor/SELinux: ✔
Minimal Seccomp whitelist: ✔

High Security Docker Microservices
- Enable User namespace
- Use app specific AppArmor if possible
- Use app specific Seccomp whitelist if possible
- Harden host system
- Restrict host access
- Consider network security

High Security runC Microservices
- Configure user namespace
- Enable AppArmor + use app specific if possible
- Use app specific Seccomp whitelist if possible
- Harden host system
- Restrict host access
- Consider network security

The Problem of Managing Secrets
Avoid environment variables/flat files
– Why? … because execve()
Use temporary secret “injection”
– Temp bind mount, load into mem-only, unmount
Vault | Keywiz

Immutable Containers also Help
Extending the basic idea of non-executable memory
pages or OpenBSD’s W^X
- data-only containers, app-only containers
Frustrating vulnerability exploitation and limiting
system post-exploitation

Networking and AuthN/AuthZ
TLS. TLS. TLS.
It’s 2016 and all network traffic should be encrypted
(and authenticated) Just do it.
Authentication/Access at the lowest possible layer. Network
access controls go against “dumb pipes” but they’re important!

Networking and AuthN/AuthZ
But I want dumb pipes!?
Then use SDN / Overlay networks on top…
Beware of cross-container networking / multi-tenant
risks with a single bridge interface

Other Security Recommendations
Have a specification / Dockumentation
Generate application-specific and overall threat models
Don’t forget about application security itself (containers and
MS can’t help if your app itself is still vulnerable)

Orchestration / Service Discovery things… Use security
Generate application-specific and overall threat models
Don’t forget about application security itself (containers and
MS can’t help if your app itself is still vulnerable)

Microservice logging and accountability is important, collect and
keep logs centrally (and actually look at them once in while)
Security is much easier to do if you make it part of your lifecycle
in the beginning “build it in don’t patch it on”
Checkout my whitepaper for more general container security info
and more info on many of the topics covered here.

Thank you!
Aaron.Grattafiori /at/ nccgroup dot trust
@dyn___

The Golden Ticket: Docker and High Security Microservices by Aaron Grattafiori

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to The Golden Ticket: Docker and High Security Microservices by Aaron Grattafiori (20)

More from Docker, Inc. (20)

Recently uploaded (20)

The Golden Ticket: Docker and High Security Microservices by Aaron Grattafiori

Editor's Notes