● Critical · CVSS 9.8

How to Fix CVE-2026-34159: llama.cpp: Unauthenticated RCE via GRAPH_COMPUTE buffer=0 bypass in llama.cpp RPC backend

Q: Is CVE-2026-34159 being actively exploited?

Not at the time of writing. It is not listed in CISA's Known Exploited Vulnerabilities catalog. That status can change, so monitor the vendor advisory and the KEV catalog if the system is exposed.

Q: How severe is CVE-2026-34159?

CVSS rates it 9.8 (Critical). Use that score to set your patch priority alongside the other items in your queue.

Q: Do I have to take llama.cpp offline to apply the patch?

It depends on the deployment. High-availability or clustered installs can usually patch one node at a time with no full outage. Standalone installs typically need a short restart. Always follow the vendor's documented upgrade steps.

Q: What if my vulnerability scanner still flags CVE-2026-34159 after I patch?

Re-run the scan after a service restart, then confirm the scanner's plugin set is up to date. Some scanners detect by banner version only and lag the official fix metadata by a release.

Other vulnerabilities in the same area that are worth patching alongside this one:

How to Fix CVE-2026-8177: Out-of-bounds Read in XML::LibXML — Out-of-bounds Read in XML::LibXML
How to Fix CVE-2026-28528: Path Traversal in BTstack — Path Traversal in BTstack
How to Fix CVE-2026-8350: Access Control Bypass in Concrete CMS , Access Control Bypass in Concrete CMS
How to Fix CVE-2026-5526: Tenda 4G03 Pro httpd access control in 4G03 Pro , Tenda 4G03 Pro httpd access control in 4G03 Pro
How to Fix CVE-2026-6867: CWE-1325: Improperly Controlled Sequential Memory Allocation in Wireshark , CWE-1325: Improperly Controlled Sequential Memory Allocation in Wireshark

*By Sai Kiran Pandrala*

Last verified: 2026-05-25

⚡ At a glance


Severity	CVSS 9.8, Critical
Actively exploited?	No
Affected	Ggml-org llama.cpp (< b8492)
Fixed in	See vendor advisory
Type (CWE)	CWE-119: CWE-119: Improper Restriction of Operations within the Bounds of a Memory Buffer

CVE-2026-34159 is a llama.cpp: unauthenticated rce via graph_compute buffer=0 bypass in llama.cpp rpc backend in Ggml-org llama.cpp. The fix is to upgrade to see vendor advisory and apply the runnable commands below.

What is CVE-2026-34159?

llama.cpp is an inference of several LLM models in C/C++. Prior to version b8492, the RPC backend's deserialize_tensor() skips all bounds validation when a tensor's buffer field is 0. An unauthenticated attacker can read and write arbitrary process memory via crafted GRAPH_COMPUTE messages.

In practical terms, a successful attacker gets remote code execution on the affected system. There is no confirmed in-the-wild exploitation listed in CISA's KEV catalog at the time of writing, but the CVSS rating still warrants prompt patching.

Am I affected?

You are affected if you run Ggml-org llama.cpp at a version listed in the Affected row above. Probe your installed build with the commands below.


# Confirm the installed version via your package manager
dpkg -l | grep -i llama.cpp   # Debian/Ubuntu
rpm -qa | grep -i llama.cpp   # RHEL/CentOS/Rocky

How to fix CVE-2026-34159

The primary fix is to upgrade to the patched build listed in the Fixed in row above (See vendor advisory). Pick the platform that matches your install and run the commands below.

Linux (Ubuntu / Debian)


# Vendor advisory: https://github.com/ggml-org/llama.cpp/security/advisories/GHSA-j8rj-fmpv-wcxw
sudo apt-get update
sudo apt-get install --only-upgrade llama-cpp
# Confirm the installed version meets or exceeds <patched-version>
dpkg -s llama-cpp | grep ^Version

Linux (RHEL / CentOS / Rocky)


sudo dnf upgrade --security llama-cpp -y
rpm -q llama-cpp

Windows (PowerShell, admin)


# Vendor advisory: https://github.com/ggml-org/llama.cpp/security/advisories/GHSA-j8rj-fmpv-wcxw
winget upgrade --id 'Ggmlorg.llamacpp' --silent --accept-source-agreements --accept-package-agreements
# If winget doesn't know the product, download the patched installer from the vendor and:
Start-Process -FilePath "$env:TEMP\llamacpp-<patched-version>.msi" -ArgumentList '/qn /norestart' -Wait

PowerShell script (Windows): detect, back up, upgrade, verify, log


# Vendor advisory: https://github.com/ggml-org/llama.cpp/security/advisories/GHSA-j8rj-fmpv-wcxw
# Run as Administrator
$ErrorActionPreference = 'Stop'
$log = "$env:ProgramData\llamacpp-Patch-CVE-2026-34159.log"
function Write-Log($msg) { "$(Get-Date -Format s)  $msg" | Tee-Object -FilePath $log -Append }

Write-Log "Starting CVE-2026-34159 remediation for Ggml-org llama.cpp"

# 1. Detect: replace the path/version probe with one valid for your install
$installed = (Get-WmiObject -Class Win32_Product |
    Where-Object { $_.Name -like '*llama.cpp*' } |
    Select-Object -First 1 -ExpandProperty Version)
Write-Log "Detected version: $installed"

if (-not $installed) {
    Write-Log "Product not installed on this host; nothing to do."
    return
}
if ([version]$installed -ge [version]'<patched-version>') {
    Write-Log "Already at fixed version $installed; no action needed."
    return
}

# 2. Backup configuration to a timestamped folder
$backup = "$env:ProgramData\llamacpp-Backup-$(Get-Date -Format yyyyMMdd-HHmm)"
New-Item -ItemType Directory -Path $backup -Force | Out-Null
$src = "$env:ProgramFiles\Ggml-org\llama.cpp"
if (Test-Path $src) { Copy-Item -Path $src -Destination $backup -Recurse -Force }
Write-Log "Backed up config to $backup"

# 3. Apply the patched installer
$installer = "$env:TEMP\llamacpp-<patched-version>.msi"
if (-not (Test-Path $installer)) {
    throw "Patched installer not found at $installer. Stage it from your software repo first."
}
Start-Process msiexec.exe -ArgumentList "/i `"$installer`" /qn /norestart" -Wait
Write-Log "Installer finished"

# 4. Verify
$verify = (Get-WmiObject -Class Win32_Product |
    Where-Object { $_.Name -like '*llama.cpp*' } |
    Select-Object -First 1 -ExpandProperty Version)
if ([version]$verify -ge [version]'<patched-version>') {
    Write-Log "SUCCESS: now at $verify (>= <patched-version>)"
} else {
    Write-Log "FAILURE: still at $verify after install"
    exit 1
}

Bash script (Linux): detect, back up, upgrade, verify, log


# Vendor advisory: https://github.com/ggml-org/llama.cpp/security/advisories/GHSA-j8rj-fmpv-wcxw
#!/usr/bin/env bash
set -euo pipefail
LOG=/var/log/llama-cpp-patch-cve-2026-34159.log
log()  { echo "$(date -Iseconds)  $*" | tee -a "$LOG"; }

log "Starting CVE-2026-34159 remediation for Ggml-org llama.cpp"

# 1. Detect installed version (works for deb and rpm packages)
if command -v dpkg >/dev/null && dpkg -s llama-cpp >/dev/null 2>&1; then
    CURRENT=$(dpkg-query -W -f='${Version}' llama-cpp)
    PKG_MGR=apt
elif command -v rpm >/dev/null && rpm -q llama-cpp >/dev/null 2>&1; then
    CURRENT=$(rpm -q --queryformat '%{VERSION}' llama-cpp)
    PKG_MGR=dnf
else
    log "llama-cpp not installed via apt or rpm; check your package manager or vendor instructions."
    exit 0
fi
log "Detected: llama-cpp=$CURRENT (manager=$PKG_MGR)"

# 2. Backup config
BACKUP=/var/backups/llama-cpp-$(date +%Y%m%d-%H%M)
mkdir -p "$BACKUP"
for d in /etc/llama-cpp /etc/${pkg%%-*} ; do
    [ -d "$d" ] && cp -a "$d" "$BACKUP/" && log "Backed up $d to $BACKUP"
done

# 3. Upgrade
if [ "$PKG_MGR" = apt ]; then
    sudo apt-get update -y
    sudo apt-get install --only-upgrade -y llama-cpp
else
    sudo dnf upgrade --security -y llama-cpp
fi

# 4. Verify
if [ "$PKG_MGR" = apt ]; then
    NEW=$(dpkg-query -W -f='${Version}' llama-cpp)
else
    NEW=$(rpm -q --queryformat '%{VERSION}' llama-cpp)
fi
log "After upgrade: $NEW"
log "Done. Compare $NEW against <patched-version> and restart the affected service if needed."

If you cannot patch immediately

These are runnable hardening commands. They reduce blast radius but they are not a replacement for the vendor patch.

Restrict the management interface to a trusted admin subnet

Linux (iptables):


sudo iptables -A INPUT -p tcp --dport 443 -s 10.0.0.0/24 -j ACCEPT
sudo iptables -A INPUT -p tcp --dport 443 -j DROP
sudo iptables-save | sudo tee /etc/iptables/rules.v4

Windows (PowerShell, admin):


New-NetFirewallRule -DisplayName 'Allow admin UI from jump network' -Direction Inbound -Action Allow -Protocol TCP -LocalPort 443 -RemoteAddress 10.0.0.0/24
New-NetFirewallRule -DisplayName 'Block admin UI from everywhere else' -Direction Inbound -Action Block -Protocol TCP -LocalPort 443

Rate-limit and log the affected endpoint at the reverse proxy


limit_req_zone $binary_remote_addr zone=admin:10m rate=10r/m;
location /<vulnerable-path> {
    limit_req zone=admin burst=5 nodelay;
    access_log /var/log/nginx/admin-access.log combined;
    proxy_pass http://backend;
}

How to verify the fix worked

Run the version probe again and confirm the running build matches the Fixed in row above.


dpkg -l | grep -i "llama.cpp"   # Debian/Ubuntu
rpm -qa | grep -i "llama.cpp"   # RHEL/CentOS/Rocky

Expected output: the package version should meet or exceed the patched build.

Re-run any vulnerability scanner you used previously and confirm the finding for CVE-2026-34159 has cleared. Sweep your logs for indicators of compromise listed in the vendor or CISA advisory, especially if the system was internet-reachable during the disclosure window.

Frequently asked questions

Is CVE-2026-34159 being actively exploited?