Skip to content
senn-techsenn-tech
Back to blog
KI & Entwicklung
AI in production2026-11-15· By Franz Senn

llama.cpp 1.0.0: three ZDI flaws on release day and a server with open defaults

On 13 November 2026 at 10:20 UTC, ggml-org tagged v1.0.0 on llama.cpp, 105 minutes after the tag on ggml (08:34 UTC). The same day the Zero Day Initiative (ZDI) published three advisories on llama.cpp, all about parsing GGUF files. The v1.0.0 release notes mention none of them. We read both and measured our standby lane on 15 Nov 2026.

In short: Semantic versioning is the scheme where the first number announces breaking changes to the public C API, the second backward-compatible additions and the third bug fixes. llama.cpp and ggml adopt it together with v1.0.0. For operators, a major bump is now a build-and-test job. The day's security fixes are already in older builds; the server hardening is new.

The three ZDI advisories

All three appeared on 13 Nov 2026 and share the attack type "local, user interaction required": the victim has to open or load a crafted file. ZDI names code execution in the context of the process as the result.

AdvisoryCVECVSSFlawFix per ZDIReported
ZDI-26-976CVE-2026-1269757.8heap buffer overflow when parsing GGUFPR 3133628 Oct 2026
ZDI-26-975CVE-2026-1269747.8integer overflow when parsing GGUFPR 3133628 Oct 2026
ZDI-26-954CVE-2026-1269767.0integer overflow in the mtmd audio projector GGUFcommit e2e7a9b18 Jun 2026

Unit: as stated on the ZDI pages of 13 Nov 2026, read on 15 Nov 2026. The CVE mapping is on the ZDI pages; CVE.org still returned 404 for all three IDs when we queried it on 15 Nov.

What we checked ourselves is in the git history of ggml-org/llama.cpp (GitHub API, 15 Nov 2026):

  • PR 31336 was merged on 28 Oct 2026 at 22:15 UTC (commit cf24317). Title: "common, mtmd: bound control vector layer indices and M-RoPE position loops". Its text describes two out-of-bounds writes and mentions neither a CVE nor ZDI.
  • The commit named for ZDI-26-954, e2e7a9b, is titled "mtmd: several bug fixes (#24784)" and dated 19 Jun 2026.
  • Both commits are ancestors of v1.0.0 and of b12282, the build of our standby lane. The v1.0.0 release notes do not list PR 31336.

Limits: that the code of the two changes fully closes each of the three ZDI descriptions rests on ZDI's note. We ran no test. PR 31336 describes control vectors and M-RoPE loops, while ZDI-26-975 and -976 speak of GGUF parsing in general. Anyone loading GGUF files from third parties should not rely on that mapping alone.

What else is in v1.0.0

The release text is 29,142 characters long (our count via the GitHub API) and has seven sections: API changes, New models, Core changes, Multi-modality changes, Server changes, UI changes, ggml changes. Eight items matter for running a lane.

Unit: PR numbers and section, read from the release text of 13 Nov 2026, as of 15 Nov 2026.

ChangeSectionPRWhat it means in operation
35 deprecated functions removed from llama.h, plus llama_set_warmup and --defrag-tholdAPI#30565Anyone compiling against libllama must rebuild. Container users notice nothing.
NextN API moves from llama-ext.h to llama.hAPI#30565Speculative decoding joins the stable surface; our lane uses MTP as draft
moe_cache_size and moe_cache_cpu_share in the contextAPI and Core#30966Expert offload to RAM becomes configurable, with a CPU share for misses
FP8 KV cache via -ctk/-ctv f8_e4m3Core#28392Less memory per context token; our lane runs q8_0
security: --cors-origins, DNS rebinding, CORS-proxy SSRF, child instancesServer#32098, #32115, #32120See the server section
security: opt-ins --allow-unauthenticated-tools, --media-allow-private-urlsServer#32113, #32115Risky conveniences sit behind a flag
/v1/audio/speech, OpenAI-compatibleServer#31763Speech output in the server itself
Web UI with chat groups and a markdown editorUI#31936, #32012, #31955No effect for us, the lane serves no UI

Also new are three text architectures (Kimi-K3.1-Pro, Ripple-65B-Preview, xing4_0), drama-3 as a TTS architecture and two vision additions (MageVL-Cowork, Kimi-K3.1). The team's Hugging Face post claims more than 150 architecture variants and all mainstream speculative-decoding methods; that is a vendor statement and we did not count. For the voice-AI starting point before 1.0.0, see our post on LFM3-Audio in llama.cpp.

Server hardening: what the maintainers say

The four server PRs were merged between 11 Nov 20:17 UTC (PR 32098) and 12 Nov 11:33 UTC (PR 32115), after b12282 appeared (11 Nov, 03:31 UTC). Our build does not contain them.

The maintainers classify them themselves. PR 32113 says that enabling tools or MCP with a non-loopback --host and no API key is now refused unless --allow-unauthenticated-tools is set. It adds, in its own words, that none of these changes is actually a vulnerability, but that the UX had to take a hit because people nitpick this. PR 32115 says the CORS proxy and media URLs could reach any local address before, by design, because users were expected to sandbox llama-server themselves. PR 32098 describes behaviour for any --cors-origin value other than *: foreign or null origins are rejected, and so are cross-site requests without an Origin header.

So the behaviour of our lane (table below) is the pre-1.0.0 default and no defect. Whether a lane on 1.0.0 with default values (--cors-origin *, no key) answers differently, we did not test; the PR texts describe changes for set values and for tools.

Measured: our standby lane

The CUDA reserve lane on our reserve inference host runs as a container from the official CUDA server image of ggml-org. It is the fallback lane; production load runs on another host. Measured on 15 Nov 2026 between 15:45 and 16:00 UTC, read-only: container inspection on the host and HTTP probes from three internal networks, with no completion request.

Unit: HTTP status and start arguments, location reserve inference host, date 15 Nov 2026, our own measurement.

What we checkedMeasuredHow
Build of the running laneb12282-aa75ff323/props
Imageofficial server-cuda image from 11 Nov 2026docker images
Distance to v1.0.077 build numbers (b12359 minus b12282, derived; not 77 published releases)tag numbers
Start argumentsbound to all interfaces, no --api-key, no tool flagsConfig.Cmd
/props, /slots without a key200HTTP probe
Foreign originechoed in the CORS header, credentials allowedprobe with Origin header
Preflight (OPTIONS)200, no Access-Control-Allow-Private-Networkside-effect-free probe
Foreign Host header200 on /healthprobe with Host: rebind.example
Reachable without a key from three internal networks200 from two, timeout from the thirdcurl, 6 s
Web UI415 on / with Accept: text/htmlprobe
KV and context-ctk q8_0 -ctv q8_0, 262,144 tokens, MTP draftConfig.Cmd

Three findings. First, without a key anyone with network access can send requests and occupy the reserve's slots. From two of the three internal networks we tested, we got through with nothing at all. That is homework for us; the project left the default open on purpose. Second, the server echoes any foreign origin and allows credentials. Whether a browser then sends requests from the public internet into an internal network is governed by the Private Network Access draft (26 Sep 2024, Draft Community Group Report), which requires an opt-in header for that case, and our lane does not send it. We tested that in no browser. Third, the Host header is not checked, which is the precondition for DNS rebinding, the thing PR 32098 addresses.

We did not probe the lane from the internet and make no statement about it.

llama.cpp builds published per day08 Nov15 · 1509 Nov12 · 1210 Nov20 · 2011 Nov14 · 1412 Nov19 · 1913 Nov20 · 20, incl. v1.0.014 Nov12 · 12022
Our own count from the github.com release list on 15 Nov 2026 around 15:55 UTC, pages 1 to 15 (150 entries, oldest b12184). Days in UTC; only fully covered days 8 to 14 Nov are counted, 7 Nov is cut off and 15 Nov is still running (29 builds so far). 11 Nov is the day our lane's image was built, v1.0.0 appeared on 13 Nov. All releases are counted, including v1.0.0. (Quelle: GitHub API ggml-org/llama.cpp, releases, own count)

New versioning and our calendar

The rules sit in a team discussion and in the Hugging Face post. They are short and answer how often you have to catch up.

Change typeNumberExample from 1.0.0
Break in the public C API in include/MAJORremoval of the 35 deprecated functions
Backward-compatible feature, backend, model supportMINORnew architecture, new flag
Bug fix with no API changePATCHserver fixes

The discussion adds that the v1.0.0 release tags are created only when llama.cpp's internal ggml copy is identical to the released ggml; the b tags remain the rolling development builds. For Docker users that means: next to the floating tag server-cuda, ghcr.io (queried 15 Nov 2026) carries the fixed tag server-cuda-v1.0.0-b12359-b1b770d. Our compose file for the reserve uses the floating tag with no digest. At the next restart it would pull a newer image than the one from 11 Nov without anyone choosing it. Pinning a tag or digest is the only way to avoid that.

What we did, and what is still open

The measurement above was read-only. Afterwards we closed the gap that bothered us most: on 15 November 2026 at about 22:10 UTC we set --api-key through the LLAMA_API_KEY environment variable and recreated the container (35 seconds to healthy). The gateway in front of the lane got the same key; before that it held a placeholder. Measured again from the same network as before: 401 without a key and with the old placeholder, /health still 200, both fallback routes of the gateway answer. Image, binding and KV settings are unchanged; the other rows of the table are not done.

OptionEffortBenefitPrice
Pin the image to a fixed tag or digestone line in composereproducible state, hardening includedrestart, throughput run afterwards
Add --api-key (done on 15 November 2026)one flagstrangers on the network no longer get throughevery client needs the key
--host to the gateway addressone flaglane reachable by the gateway onlyrework once a second consumer appears
Test FP8 for KVmeasurement seriesmore context on the same cardprove quality with real requests first
Test the MoE expert cachemeasurement serieshope for large mixtures on little VRAMthe lane runs a dense model, not a candidate

FP8 would halve the bit width of q8_0 and so roughly halve KV memory; that is derived from bit width and not measured by us. How we used FP8 in the KV cache on our production path is described in Flash-Next in production.

Open points

  • Whether the code of PR 31336 and commit e2e7a9b closes each ZDI description: ZDI's statement only, no test.
  • How a 1.0.0 server with default values answers a foreign origin and a foreign Host header: not measured.
  • CVE records: CVE.org returned 404 for all three IDs on 15 Nov, NVD had no record for CVE-2026-126975, and the queries for the other two gave no usable result. The IDs therefore rest on the ZDI pages alone.
  • Origin of the two GGUF files on our lane: not part of this measurement.
  • Reachability from the internet: not checked.
  • Browser behaviour for Private Network Access: not tested; the draft dates from 2024.

Read next: the ImageMagick policy against ImageTragick2 · Huawei MACaQ for GGUF, measured · the DeepSeek API: speech recognition, speech output, OCR

Further Reading

Primary sources (read 15 Nov 2026)

Security and outside sources

From us

Questions?
Do the three ZDI flaws affect our running llama.cpp lane?+

Not any more, going by git history: pull request 31336, which ZDI names as the fix (merged 28 Oct 2026), and the commit e2e7a9b named for ZDI-26-954 are both ancestors of build b12282, the one our standby lane runs. That is derived from commit history; we ran no exploit. We did not check whether the code of the two changes covers every ZDI description.

Should we rebuild our inference lane right away because of 1.0.0?+

Not because of the three flaws if your build is b12282 or newer. Yes for the server hardening if the server is reachable from other networks or tools are enabled. A restart costs a maintenance window and calls for a throughput comparison. The release date alone is no reason.

How do I tell whether my own installation is affected?+

Three quick checks. Open /props and read the build number. Check the start arguments for --host and --api-key; --host 0.0.0.0 without a key means open to anyone with network access. Send a request with a foreign Origin header and look at what Access-Control-Allow-Origin returns. Also work out from git history whether your build contains commit cf24317.