--- title: "NVIDIA ForgePtr CVE-2026-47501: GPU Container to Root" url: https://playciso.com/blog/nvidia-forgeptr-cve-2026-47501-gpu-container-escape updated: 2026-11-19 section: Blog publisher: PlayCISO --- # NVIDIA ForgePtr CVE-2026-47501: GPU Container to Root > ForgePtr (CVE-2026-47501) in NVIDIA's Linux GPU driver lets an unprivileged container with GPU access get host root. Fixed versions, what to patch first. Source: https://playciso.com/blog/nvidia-forgeptr-cve-2026-47501-gpu-container-escape · Published: 2026-11-19 · Publisher: PlayCISO (https://playciso.com) Primary source: https://nvidia.custhelp.com/app/answers/detail/a_id/5861/~/security-bulletin%3A-nvidia-gpu-display-driver---september-2026 --- **CVE-2026-47501**, which researchers at [Stealthium](https://stealthium.io/blog/forgeptr) named **ForgePtr**, is a flaw in the kernel-mode layer of **NVIDIA's GPU display driver for Linux**. [NVIDIA's September 2026 security bulletin](https://nvidia.custhelp.com/app/answers/detail/a_id/5861/~/security-bulletin%3A-nvidia-gpu-display-driver---september-2026), published on **30 September 2026**, rates it CVSS 3.1 **7.8 (High)** and lists fixed drivers on every supported Linux branch. In a write-up published in November, Stealthium says an **unprivileged process inside a container with only GPU access can use it to become root on the host**. Stealthium released no proof of concept, we found no report of exploitation in the wild, and the CVE is **not in CISA's KEV catalog** as of 19 November 2026. **Who this is for:** anyone who runs NVIDIA GPUs on Linux and lets more than one user, team, customer or container share a node: AI and ML platforms, Kubernetes GPU pools, research clusters, GPU cloud providers and the people who buy capacity from them. *Based on NVIDIA's bulletin (read in full, including its version table and acknowledgements) and Stealthium's write-up, both read on 19 November 2026. PlayCISO is not affiliated with NVIDIA or Stealthium.* **Confirmed vs not yet known (19 Nov 2026)** - Confirmed by NVIDIA: CVE-2026-47501 is an out-of-bounds write (CWE-787) in the Linux kernel-mode layer, triggered by "supplying mismatched memory buffers during event buffer setup"; CVSS 7.8, local, low privileges, no user interaction; fixed versions on R580, R595, R610 and R615 and in vGPU 19.6 and 20.2. - Stealthium's claims (not restated by NVIDIA): a working root exploit from an unprivileged account; container-to-host escape through core_pattern; disabling ASLR, kernel lockdown and module signature enforcement on a Secure Boot machine. Stealthium says it verified every step on affected hardware. - Not reported: exploitation in the wild. Not in CISA KEV on 19 November. No public proof of concept. - Sources disagree: NVIDIA lists 610.57.04 as the fixed R610 driver; Stealthium says the flaw is present through 610.57.04 and that it got root on it. See below. - Not known: whether the proprietary (closed) kernel modules behave the same way in practice; Stealthium's view that they are affected is an inference. ## What ForgePtr is, in plain words Stealthium says the bug sits in the driver's event-buffer setup path: a handle supplied by the caller is converted into a kernel pointer, the driver never checks whether that conversion worked, keeps the raw value and reports success. When the object is later freed, the kernel decrements a counter at an address the attacker chose. Repeating that lets an attacker write chosen bytes to a chosen kernel address. Stealthium says an attacker also needs a separate unprivileged way to learn kernel addresses (a KASLR leak), and nothing else: an ordinary local account, with no special group membership and no `CAP_SYS_ADMIN`. From there, Stealthium describes rewriting `core_pattern`, the kernel setting that says which program to run when a process crashes. Because `core_pattern` is global kernel state and not namespaced, a process inside a container can point it at a program in its own filesystem; when a child process crashes, the kernel runs that program as real host root. Stealthium calls the result "total host takeover": every other tenant's data on the node readable and writable, and persistence on the node. It says the same write primitive can switch off ASLR, kernel lockdown and module signature enforcement, and load an unsigned kernel module on a Secure Boot machine. The point for leaders is simple: **on a shared GPU node, the GPU driver is part of your isolation boundary.** Container namespaces, seccomp and Pod Security settings do not help if the driver every container talks to will write wherever it is told. ## Affected and fixed versions From NVIDIA's bulletin (Linux rows for CVE-2026-47501; "All driver versions prior to" the fixed version are affected): | Product | Branch | Fixed version | | --- | --- | --- | | GeForce, NVIDIA RTX / Quadro / NVS, Tesla (data center) | R580 | 580.178.04 | | GeForce, NVIDIA RTX / Quadro / NVS, Tesla | R595 | 595.91.07 | | GeForce, NVIDIA RTX / Quadro / NVS, Tesla | R610 | 610.57.04 (disputed, see below) | | GeForce, NVIDIA RTX / Quadro / NVS, Tesla | R615 | 615.71.09 | | vGPU guest driver, Linux | vGPU 19 / vGPU 20 | 580.178.04 (vGPU 19.6) / 595.91.07 (vGPU 20.2) | | Virtual GPU Manager (XenServer, VMware vSphere, RHEL KVM, Ubuntu) | vGPU 19 / vGPU 20 | 580.178.05 (vGPU 19.6) / 595.91.04 (vGPU 20.2) | *Source: [NVIDIA bulletin 5861](https://nvidia.custhelp.com/app/answers/detail/a_id/5861/~/security-bulletin%3A-nvidia-gpu-display-driver---september-2026). The same table also lists Windows vGPU guest driver and Virtual GPU Manager rows for this CVE (vGPU 19.6 and 20.2); check the bulletin for those. NVIDIA says the table "might not be a comprehensive list" and that earlier branches not listed may also be affected: if you run one, upgrade to the latest branch release. vGPU and Cloud Gaming updates come from the NVIDIA Licensing Portal.* ### The R610 disagreement Stealthium's write-up says the flaw is "present in source in every open-source driver release from 515.43.04 (2022-05-09) through 610.57.04 (2026-08-03)" and was "runtime-confirmed to root on 610.43.02 and 610.57.04". NVIDIA's table lists 610.57.04 as the *fixed* R610 version. We cannot tell from the two documents which is right, and neither addresses the other. Stealthium's own advice is to "update to 615.71.09 or later", the version both sources agree is fixed. If you are on R610, the safe reading is to move to 615.71.09 or later, or ask NVIDIA to confirm the R610 fix before you rely on it. ### Who gets credit Stealthium says the vulnerability was found by its senior security researcher Branislav Brzak and reported to NVIDIA. NVIDIA's bulletin lists no acknowledgement for CVE-2026-47501, and notes that "All unacknowledged CVEs were found internally." The same bulletin fixes a second Linux kernel-mode flaw, CVE-2026-47591 (an unprivileged user can write to memory marked read-only, CVSS 7.8, which some coverage calls NvRoot), in the same driver versions. Stealthium also says it has "25+ further High or Critical vulnerabilities" under responsible disclosure across the AI runtime stack, so expect more GPU-stack advisories. ## Timeline | Date | Event | Source | | --- | --- | --- | | 9 May 2022 | First public release of NVIDIA's open GPU kernel modules (515.43.04); Stealthium says the flaw is already present | Stealthium | | 9 Sep 2026 | NVIDIA releases driver 615.71.09, which Stealthium says adds the missing error check | Stealthium | | 30 Sep 2026 | NVIDIA publishes its September 2026 GPU display driver bulletin (revision 1.0.0) and assigns CVE-2026-47501 | NVIDIA | | 13 Oct 2026 | Bulletin revision 2.0.0 (corrects the CVSS score of a different CVE, CVE-2026-47544) | NVIDIA | | Nov 2026 | Stealthium publishes "Introducing ForgePtr": root and container escape described, no proof of concept | Stealthium | | 18 Nov 2026 | CISA KEV catalog version 2026.11.18 does not list CVE-2026-47501 (checked 19 Nov) | CISA | ## What security leaders should do this week - Inventory every GPU host's driver version. On each node, nvidia-smi --query-gpu=driver_version --format=csv,noheader or cat /proc/driver/nvidia/version. Include Kubernetes GPU node pools, Slurm and research clusters, CI runners with GPUs, workstations shared by several users, and vGPU hosts and guests. - Patch shared nodes first. Rank by who can run code on the node: GPU nodes that run other customers' or other teams' containers, notebooks or CI jobs come first; single-owner workstations last. Moving to a fixed driver means reloading the kernel modules, so plan a drain and reboot for each node. - Check how your drivers are delivered. If drivers come from a node image (managed Kubernetes, a cloud machine image) or from a driver container (for example NVIDIA's GPU Operator), the fix arrives through that image or that setting, not through a package update on a running node. Confirm the version your provider ships, and update the image or driver version setting. - Until patched, keep untrusted workloads off. Stealthium says no sysctl, seccomp profile, IOMMU or device-isolation setting closes the flaw. Use taints and node pools to keep untrusted or multi-tenant GPU jobs off unpatched nodes, give each tenant its own node, and prefer VM-level isolation over containers for tenants you do not trust (and patch the vGPU manager too). - Watch the settings an attacker would change. Stealthium names core_pattern, kernel lockdown and sig_enforce as targets. A periodic check of /proc/sys/kernel/core_pattern, /sys/kernel/security/lockdown and /proc/sys/kernel/randomize_va_space against a known baseline on GPU nodes is cheap, and an unexpected change is worth a page. - Ask your GPU cloud. If you rent GPU capacity on shared hosts, ask the provider which driver its hosts run and when ForgePtr was patched. Put the answer in your vendor file. - Set a deadline and record it. Not in KEV and no known exploitation, but a full host takeover on shared infrastructure with a named researcher holding a working exploit. Score it, set the date, and log any exception with an owner. ## Put it to work - [SSVC calculator](/tools/ssvc-calculator): turn exploitation, exposure and impact into a patch decision you can defend. - [Vulnerability prioritizer](/tools/vulnerability-prioritizer): rank CVE-2026-47501 against the rest of your queue. - [KEV checker](/tools/kev-checker): re-check whether CVE-2026-47501 has been added to CISA's exploited list. - [Hypervisor hardening lab](/tools/hypervisor-hardening-lab): practise the host-level isolation decisions that matter when a guest or container can reach the kernel. The vendor timeline is in the [Breach Index entry for NVIDIA](/breach-index/nvidia). It is a vulnerability record, not a breach: nothing reported shows an intrusion at NVIDIA. ## FAQ ### What is ForgePtr? ForgePtr is the name Stealthium gave to CVE-2026-47501, a flaw in the kernel-mode layer of NVIDIA's GPU display driver for Linux. NVIDIA describes it as an out-of-bounds write a user can cause "by supplying mismatched memory buffers during event buffer setup", rated CVSS 3.1 7.8 (High). Stealthium says it lets an unprivileged local user write chosen bytes into kernel memory and become root. ### Can ForgePtr be used to escape a container? Stealthium says yes: an unprivileged process in a container that has only GPU access can overwrite core_pattern, a kernel setting that is not namespaced, and get a program run as root in the host's namespaces. Stealthium says it verified each step on affected hardware. NVIDIA's bulletin lists the impacts as code execution, privilege escalation, data tampering, denial of service and information disclosure, and does not discuss containers. ### Which NVIDIA driver versions fix CVE-2026-47501? NVIDIA's bulletin lists these Linux fixed versions for GeForce, NVIDIA RTX/Quadro/NVS and Tesla: 580.178.04 (R580), 595.91.07 (R595), 610.57.04 (R610) and 615.71.09 (R615). For vGPU it lists 19.6 and 20.2. Stealthium says the flaw is present through 610.57.04 and that it got root on 610.57.04, and it advises 615.71.09 or later. Until that is reconciled, 615.71.09 or later is the version both sources agree is fixed. ### Is ForgePtr being exploited? We found no report of exploitation in the wild. CVE-2026-47501 was not in the CISA Known Exploited Vulnerabilities catalog when we checked on 19 November 2026, NVIDIA's bulletin does not mention exploitation, and Stealthium says it built a working exploit but has not released a proof of concept. ### Is there a workaround for ForgePtr? Stealthium says there is no meaningful workaround short of patching: the bug is reachable by anything that can open the GPU device, so no sysctl or seccomp profile closes it, and IOMMU, device isolation and confidential-compute features do not help because the kernel itself performs the write. The interim control is to keep untrusted workloads off unpatched GPU nodes. ### Does ForgePtr affect the NVIDIA Container Toolkit? Neither NVIDIA's bulletin nor Stealthium names the NVIDIA Container Toolkit. The flaw is in the kernel driver on the host. The Toolkit and the Kubernetes device plugin matter because they are what give a container access to the GPU device, which is all Stealthium says an attacker needs. ### Is the closed-source NVIDIA driver affected too? NVIDIA's bulletin covers the "GPU Display Driver for Linux" and does not split open and proprietary kernel modules. Stealthium traced the flaw through the open GPU kernel modules back to the first public release, 515.43.04 (May 2022), and says it "likely predates open-sourcing in the closed driver too", which is its inference, not a test result. Use the bulletin's fixed versions whichever module you run. ### Was NVIDIA breached? No. ForgePtr is a vulnerability in an NVIDIA product, not an intrusion at NVIDIA. The exposure is to organisations that run affected drivers on hosts shared by users or containers they do not fully trust. ## Sources - [NVIDIA: Security Bulletin: NVIDIA GPU Display Driver, September 2026 (a_id 5861)](https://nvidia.custhelp.com/app/answers/detail/a_id/5861/~/security-bulletin%3A-nvidia-gpu-display-driver---september-2026) (initial release 30 Sep 2026, revision 2.0.0 13 Oct 2026). Primary source for the description, CVSS and fixed versions; also mirrored on [NVIDIA's product-security GitHub](https://github.com/NVIDIA/product-security/blob/main/2026/5861/5861.md). - [CVE record: CVE-2026-47501](https://www.cve.org/CVERecord?id=CVE-2026-47501) - [Stealthium: Your Configs, My Choices: Introducing ForgePtr](https://stealthium.io/blog/forgeptr) (Branislav Brzak, Chris Hosking and the Stealthium Core Team, Nov 2026). Source for the exploit chain, the affected range in the open modules and the R610 claim. - [CISA Known Exploited Vulnerabilities catalog](https://www.cisa.gov/known-exploited-vulnerabilities-catalog) (version 2026.11.18, checked 19 Nov 2026: not listed) - [NVIDIA Product Security](https://www.nvidia.com/security/) *This article is educational guidance for security and infrastructure leaders. Check NVIDIA's bulletin for revisions before acting; fixed versions can change.*