ggml-org llama.cpp Flaw CVE-2026-118375: Update Now
Published October 29, 2026 · Last updated October 29, 2026
The short version: CVE-2026-118375 is a high-severity security flaw in ggml-org llama.cpp. The National Vulnerability Database (NVD) published it on October 28, 2026, and the score it lists is 8.1 out of 10. When FraudMeNot checked on October 29, 2026, CISA had not added CVE-2026-118375 to its Known Exploited Vulnerabilities catalog, so there was no public U.S. government confirmation of attacks yet. A fix is available, which is why this page focuses on who should update and how.
What the official record says about CVE-2026-118375
The record was filed by vulncheck.com. In its own words: "llama.cpp b8841 before b11865 contains an out-of-bounds write vulnerability in the ggml-rpc RDMA transport rdma_recv() function that allows unauthenticated peers to overflow the rpc-server stack. Attackers advertising RDMA capabilities in the HELLO message can send up to 256 KiB into a 1-byte stack variable in rpc_serve_client(), crashing the server or potentially executing code."
Why llama.cpp admins should care
Severity scores use the Common Vulnerability Scoring System (CVSS). the address listed on vulncheck.com rated CVE-2026-118375 8.1 (High) under CVSS 3.1. Reading that rating in everyday terms: it can be reached over a network, including the internet; an attacker does not need an account; no one has to click or open anything. If it is abused, the rated damage to keeping data private is high, to keeping data unaltered is high, and to keeping llama.cpp running is high. Scoring string: CVSS:3.1/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H.
Scores can differ between the vendor and NVD because each judges the worst realistic case differently. For this flaw, the address listed on vulncheck.com gives it 9.2 (Critical, CVSS 4.0).
Affected llama.cpp releases
According to the record, these releases are vulnerable:
- b8841 up to, but not including, b11865
The type of bug behind CVE-2026-118375
NVD classifies it using MITRE's Common Weakness Enumeration:
- CWE-787, Out-of-bounds Write. MITRE describes this class of bug as follows: "The product writes data past the end, or before the beginning, of the intended buffer."
Vendor remediation for CVE-2026-118375
The vendor advisory remediation for operators of llama.cpp, quoted for this page, is: "Move to llama.cpp release b11865 or newer. The fix (pull request 30399, titled rpc: add bounds check in rdma_recv) makes the RDMA receive path check that each incoming completion fits in the remaining buffer before copying it, logs the error, reposts the receive slot, and drops the connection instead of corrupting memory. Only builds that run the optional rpc-server with RDMA transport are exposed, so check whether you start rpc-server at all. Rebuild or re-pull any app, Docker image, or desktop tool that bundles its own copy of ggml-rpc, because those will not update on their own. Bind rpc-server to localhost or a private, firewalled network segment, since the RPC protocol has no login of its own and anyone who can reach the port can talk to it."
What to do about it
- Update now: Move to llama.cpp release b11865 or newer. The fix (pull request 30399, titled rpc: add bounds check in rdma_recv) makes the RDMA receive path check that each incoming completion fits in the remaining buffer before copying it, logs the error, reposts the receive slot, and drops the connection instead of corrupting memory. Only builds that run the optional rpc-server with RDMA transport are exposed, so check whether you start rpc-server at all. Rebuild or re-pull any app, Docker image, or desktop tool that bundles its own copy of ggml-rpc, because those will not update on their own. Bind rpc-server to localhost or a private, firewalled network segment, since the RPC protocol has no login of its own and anyone who can reach the port can talk to it.
- Who should act: IT teams, managed service providers, and developers who run or ship llama.cpp. Check every place it runs, including servers, build systems, containers, and older machines nobody watches closely.
- Read the official advisory from ggml-org (linked under Sources) for exact fixed builds and any temporary workaround if you cannot patch today.
- After patching, confirm the new version is actually running and review logs from October 28, 2026 onward for anything unusual.
- Home users: you do not need to buy or download anything because of CVE-2026-118375. A caller, text, or pop-up claiming your computer is "infected" with this CVE and asking for money or remote access is a scam.
CVE-2026-118375 record details from CISA and NVD
- Attack complexity for CVE-2026-118375: high
- CVE-2026-118375 CVSS scope: unchanged
- NVD sub-scores for CVE-2026-118375: exploitability 2.2, impact 5.9.
- Who mapped the weakness type: the address listed on vulncheck.com (Primary)
- References in the NVD record: 6.
Dates to know
- October 28, 2026: NVD publishes CVE-2026-118375.
- October 29, 2026: FraudMeNot checks the CISA KEV catalog; CVE-2026-118375 is not listed.
- October 28, 2026: most recent NVD update (record status: Awaiting Analysis).
Official references
Links listed in the NVD record (labels in brackets come from NVD):
- github.com/ggml-org/llama.cpp
- github.com/ggml-org/llama.cpp/blob/620d2225cc436ab0b001e3200e590e3b286ef9ab/ggml/src/ggml-
- github.com/ggml-org/llama.cpp/commit/9a0fde9967fa1a35b6e5dd1f47e1a3648b6fecc1
- github.com/ggml-org/llama.cpp/pull/30399
- github.com/ggml-org/llama.cpp/releases/tag/b11865
- www.vulncheck.com/advisories/llama-cpp-b8841-before-b11865-stack-based-buffer-overflow-via
Sources
- NVD: CVE-2026-118375
- CVE.org: CVE-2026-118375
- ggml-org advisory
- ggml-org advisory
- ggml-org advisory
- MITRE CWE-787
General education only, not security advice. Facts checked against the sources above on the date shown.