llama.cpp denial of service in llama-server sampler init
Affects llama-server in llama.cpp when a request to /completion or /v1/completions includes a grammar_triggers type outside 0-3; no special setup is needed. The fix is patch 9f26fe1.
- Exploited
- No known exploitation
- Fix
- Vendor patch: 9f26fe1
- Who is exposed
- Default setups
Summary
Improper input validation in the llama-server component of ggml-org llama.cpp, in the common_sampler_init function, lets a remote attacker send manipulated grammar_triggers type input without logging in. Successful exploitation could cause a limited loss of availability.
Written by a model
Original description
A security vulnerability has been detected in ggml-org llama.cpp 0.7.0-dev. This affects the function common_sampler_init of the file common/sampling.cpp of the component llama-server. Such manipulation of the argument grammar_triggers[].type leads to improper input validation. It is possible to launch the attack remotely. The exploit has been disclosed publicly and may be used. Upgrading to version b12405 mitigates this issue. The name of the patch is 9f26fe11c73c7e74a5e99e7f1b6370ad8e1cb22d. It is advisable to upgrade the affected component.
Who should check
Teams running llama-server from ggml-org llama.cpp that accept /completion or /v1/completions requests, including builds around 0.7.0-dev.
“allows remote SIGABRT in llama-server”
What to do
Check every install: default setups are affected.
Apply the vendor's patch: 9f26fe1.
Check your systems with an agent
Paste this into a coding or ops agent. It checks your systems read-only and reports a table per asset.
Check whether my systems are affected by CVE-2026-134540 (ggml-org llama.cpp llama-server sampling.cpp common_sampler_init input validation), using the cve.now data below.
Rules: work read-only. Do not run exploit code, proof-of-concept scripts or scanners that send attack payloads, and do not change any configuration, package, image or deployment. Report findings only.
What cve.now has for CVE-2026-134540 (entry last updated 2026-11-21):
Affected:
- ggml-org llama.cpp: versions 0.7.0-dev
Vendor patches: 9f26fe1 (source: https://github.com/ggml-org/llama.cpp/commit/9f26fe11c73c7e74a5e99e7f1b6370ad8e1cb22d). Check each asset for these patch ids, not only its version.
Conditions (all must be true for the bug to be hit):
- A request to /completion or /v1/completions includes a grammar_triggers entry whose type is outside the valid range 0-3. (source: https://github.com/ggml-org/llama.cpp/issues/30835)
- Works in the default configuration; no grammar or special flag is needed. (source: https://github.com/ggml-org/llama.cpp/issues/30835)
Setup: the bug is reachable in a default setup; no special feature or setting is needed.
Attacker needs: network access, no login, no user action.
CISA KEV: CVE-2026-134540 is not listed as of cve.now's last read.
This KEV status is for CVE-2026-134540 only. If you also check another CVE, look up its own KEV status; never assume it from this one.
How to check:
- Installed software: check what is actually running, not what a repo declares. Read the version from each host, container image and Kubernetes node (the binary's version output, the OS package database, image layers), and check whether the conditions above hold there. Distribution packages may carry a backported fix under an older version number; confirm against the distribution's own advisory.
Verify against the linked advisories before concluding, and cite the date you read each one:
- https://github.com/ggml-org/llama.cpp/pull/31478
- https://github.com/ggml-org/llama.cpp/commit/9f26fe11c73c7e74a5e99e7f1b6370ad8e1cb22d
- https://github.com/ggml-org/llama.cpp/releases/tag/b12405
- https://github.com/ggml-org/llama.cpp/issues/30835
- https://vuldb.com/vuln/441681
- https://vuldb.com/vuln/441681/cti
- https://vuldb.com/cve/CVE-2026-134540
- https://vuldb.com/submit/1052496
Do not conclude an asset is unaffected unless you checked it and the evidence covers both its version and the conditions above. If you could not check something, mark it unknown and say what is missing.
Report a table with one row per asset and these columns: Asset | Version or build | Status (affected, fixed, not applicable, unknown) | Compensating controls | Exposure (internet-facing, internal, none) | Evidence | Next action.
Entry: https://cve.now/CVE-2026-134540
For follow-up questions, use the cve.now MCP server at https://mcp.cve.now/mcp (no authentication), for example execute with name cve.get and arguments {"id": "CVE-2026-134540"}.Affected versions
- ggml-org llama.cpp0.7.0-dev
cve.now Index
- Exploitation?/40Unknown. No CISA SSVC exploitation rating yet
- Exposure22/25Default setups, reachable over the network without a login
- Reach?/20Unknown. No tracked technology or package matched
- Severity6.9/10CVSS 6.9
- No patch0/5No signal
What would change the score
- Would score about 69 if added to CISA KEV
- Would score about 64 if CISA's SSVC rates exploitation active
Timeline
- CVE published
- First seen by cve.now
Evidence 12
| Kind | Source |
|---|---|
| CVE record 2Establishes: The identifier and description. | |
| Record | CVE List · CVE-2026-134540 |
| Record | NVD · CVE-2026-134540 |
| Public exploit code 1Establishes: Public exploit code exists (Exploit-DB, Metasploit, Nuclei and GitHub). | |
| Exploit | gist.github.com · Yunzez/90b17f91d7f1955805e0045a1a76fb6d#file-repro_llamacpp_grammar_triggers_abort-py |
| Other links 9 | |
| Advisory | VulDB · Advisory #441681 |
| Advisory | VulDB · Threat intelligence #441681 |
| Advisory | VulDB · CVE-2026-134540 record |
| Advisory | VulDB · Submission #1052496 |
| Patch | GitHub · ggml-org/llama.cpp · Pull request #31478 |
| Patch | GitHub · ggml-org/llama.cpp · Commit 9f26fe11 |
| Patch | GitHub · ggml-org/llama.cpp · Release b12405 |
| Report | GitHub · ggml-org/llama.cpp · Issue #30835 |
| Repository | GitHub · ggml-org/llama.cpp · Repository |
Questions about CVE-2026-134540
Is CVE-2026-134540 being exploited?
It is not in CISA's Known Exploited Vulnerabilities catalog, and CISA's SSVC assessment does not mark exploitation as active. That does not mean it has not been attacked: vendors and researchers can report attacks before CISA lists them. No public exploit code is tracked.
Source: cve.now feed data
What does this mean?
A request to /completion or /v1/completions with a grammar_triggers entry whose type is outside the range 0-3 makes llama-server abort with SIGABRT. The process crashes, taking down every slot and client.
Source: github.com, github.com
Which versions are affected by CVE-2026-134540?
llama.cpp: version 0.7.0-dev.
Source: cve.now feed data
Am I affected, and how do I check?
The facts do not list affected version ranges. The bug works in the default configuration and needs no grammar or special flag; the trigger is an out-of-range grammar_triggers type in a completion request. To check a build, see whether it includes patch 9f26fe1.
Source: github.com, github.com
How do I fix CVE-2026-134540?
Apply patch 9f26fe1, which makes llama.cpp reject out-of-range grammar_triggers types instead of aborting. No fixed version number is named in the facts.
Source: github.com, github.com
Related
- CVE-2026-118375llama.cpp b8841 before b11865 Stack-based Buffer Overflow via ggml-rpc RDMA rdma_recv()
- CVE-2026-107183llama.cpp before b11393 Use-After-Free via common_chat_peg_mapper chat_parser
- CVE-2026-43632llama.cpp b7492–b9060 Use-After-Free in Tokenization Endpoints
- CVE-2026-43631llama.cpp b7492–b9060 Use-After-Free RCE via llama-server
- CVE-2026-43629llama.cpp b4882–b9058 Buffer Overflow in KV Cache State Restore
- CVE-2024-58400llama.cpp before b3701 Out-of-Bounds Read in llama_decode via Token IDs
No corrections so far. Something wrong on this page? Report an error.