Cyber Resilience

CVE-2026-21869

Memory Safety in Ggml Llama.Cpp

Public PoCMemory Safety
Published
08 January 2026
Modified
02 February 2026
Patch / advisory
View advisory
CVSS Score v3.1 8.8
Click a component to see what it means
Raw vectorCVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H
EPSS Score 0.0045 37th percentile
Risk Priority 64 floored blend · peak EPSS

Summary

CVE-2026-21869 is a high-severity Out-of-bounds Write (CWE-787) vulnerability in Ggml Llama.Cpp. Its CVSS base score is 8.8 (High).

Operationally, exploitation aligns with the MITRE ATT&CK technique Exploitation for Privilege Escalation (T1068); ranked at the 37th percentile by exploit likelihood (below the median); it is not currently listed in the CISA KEV catalog; a public proof-of-concept is referenced.

This vulnerability is AI-related — categorised as NLP and Transformers; in the LLM/Generative AI Risks risk domain.

The strongest mitigations our analysis identified map to SA-11 (Developer Testing and Evaluation) and SA-15 (Development Process, Standards, and Tools) — see the control section below for these in your framework.

Deeper analysis AI-assisted summary

Synthesised by an AI model from the NVD description and linked references — a reading aid, not an authoritative source.

CVE-2026-21869 is a memory corruption vulnerability in llama.cpp, a C/C++ inference engine for large language models (LLMs). In commits up to 55d4206c8, the server's completion endpoints parse the n_discard parameter directly from JSON input without validating that it is non-negative. Supplying a negative value, combined with a full context, results in a reversed range and negative offset passed to llama_memory_seq_rm/add functions, triggering out-of-bounds memory writes during the token evaluation loop. This issue is classified under CWE-787 (Out-of-bounds Write) with a CVSS v3.1 base score of 8.8 (AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H).

Remote attackers can exploit this vulnerability by sending crafted JSON requests to the llama.cpp server's completion endpoints, requiring minimal user interaction such as submitting the malicious input. No authentication or privileges are needed, making it accessible over the network with low complexity. Successful exploitation leads to deterministic memory corruption, which can crash the server process or potentially enable remote code execution (RCE) by overwriting critical memory regions.

The GitHub Security Advisory (GHSA-8947-pfff-2f3c) details the issue but notes there is no fix available at the time of publication on January 8, 2026. Security practitioners should monitor the llama.cpp repository for patches, avoid exposing the server publicly, and validate all JSON inputs server-side until remediation is released.

This vulnerability is particularly relevant to AI/ML deployments, as llama.cpp is widely used for efficient LLM inference, potentially exposing model serving infrastructure to compromise. No real-world exploitation has been reported in available data.

EU & UK References

Vulnerability Data

llama.cpp is an inference of several LLM models in C/C++. In commits 55d4206c8 and prior, the n_discard parameter is parsed directly from JSON input in the llama.cpp server's completion endpoints without validation to ensure it's non-negative. When a negative value…

more

is supplied and the context fills up, llama_memory_seq_rm/add receives a reversed range and negative offset, causing out-of-bounds memory writes in the token evaluation loop. This deterministic memory corruption can crash the process or enable remote code execution (RCE). There is no fix at the time of publication.

CWE(s)
CWE-787

AI Security AnalysisAI

AI Category
NLP and Transformers
Risk Domain
LLM/Generative AI Risks
OWASP Top 10 for LLMs 2025
None mapped
Classification Reason
Matched keywords: llama.cpp, llm

Related Threats

MITRE ATT&CK Enterprise Techniques

T1068 Exploitation for Privilege Escalation Privilege Escalation
Adversaries may exploit software vulnerabilities in an attempt to elevate privileges.
attack.mitre.org →
T1190 Exploit Public-Facing Application Initial Access
Adversaries may attempt to exploit a weakness in an Internet-facing host or system to initially access a network.
attack.mitre.org →
T1203 Exploitation for Client Execution Execution
Adversaries may exploit software vulnerabilities in client applications to execute code.
attack.mitre.org →
T1210 Exploitation of Remote Services Lateral Movement
Adversaries may exploit remote services to gain unauthorized access to internal systems once inside of a network.
attack.mitre.org →
T1211 Exploitation for Stealth Stealth
Adversaries may exploit vulnerabilities to evade detection by hiding activity, suppressing logging, or operating within trusted or unmonitored components.
attack.mitre.org →
T1212 Exploitation for Credential Access Credential Access
Adversaries may exploit software vulnerabilities in an attempt to collect credentials.
attack.mitre.org →
Derived from this CVE’s CWE(s) via the direct CWE→ATT&CK cross-walk.

CVEs Like This One

CVE-2024-23496Same product: Ggml Llama.Cpp
CVE-2024-21836Same product: Ggml Llama.Cpp
CVE-2024-42479Same product: Ggml Llama.Cpp
CVE-2024-21825Same product: Ggml Llama.Cpp
CVE-2024-23605Same product: Ggml Llama.Cpp
CVE-2024-21802Same product: Ggml Llama.Cpp
CVE-2020-0986Shared CWE-787
CVE-2023-44807Shared CWE-787
CVE-2023-33635Shared CWE-787
CVE-2021-30761Shared CWE-787

Affected Assets

ggml
llama.cpp
all versions

Mitigating Controls

Mitigating Controls (NIST 800-53 r5) AI

SA-11 Developer Testing and Evaluation good match

Developer testing and evaluation (including fuzzing and bounds checks) finds out-of-bounds write flaws before deployment.

SA-15 Development Process, Standards, and Tools partial match

Requiring documented secure-development standards and tools can mandate bounds-checked coding practices that avoid the weakness.

SI-10 Information Input Validation partial match

Input validation can structurally reject or sanitize data that would otherwise trigger an out-of-bounds write.

SI-16 Memory Protection partial match

Memory-protection mechanisms limit the exploitability and blast radius of a successful out-of-bounds write.

Mitigating Controls (NIST CSF 2.0) AI

Derived directly from the weakness types (CWEs) cited in the NVD entry via our AI-authored CWE→CSF cross-walk (authority under review) — links open the control.

PR.PS-06 mostly match
prevents

Secure-development practices (static analysis, bounds checking, code review) are the primary means of preventing out-of-bounds writes.

ID.RA-01 partial match
prevents

Vulnerability scanning and recording can discover out-of-bounds write flaws so they can be remediated.

PR.PS-02 partial match
prevents

Patching or replacing vulnerable software directly eliminates known instances of this coding weakness.

Mitigating Controls (ISO/IEC 27001:2022 Annex A) AI

Derived directly from the weakness types (CWEs) cited in the NVD entry via our AI-authored CWE→ISO cross-walk (authority under review) — links open the control.

A.8.29 Security testing in development and acceptance mostly match
finds

Security testing in development and acceptance can detect and prevent out-of-bounds write defects.

A.8.25 Secure development life cycle partial match
prevents

Secure development life cycle mandates practices that prevent out-of-bounds writes.

A.8.26 Application security requirements partial match
prevents

Application security requirements can specify bounds-checking and safe memory handling.

A.8.27 Secure system architecture and engineering principles partial match
prevents

Secure architecture and engineering principles reduce the likelihood of buffer overflows.

A.8.28 Secure coding partial match
prevents

Secure coding directly addresses out-of-bounds writes through language choice and coding standards.

A.8.32 Change management partial match
prevents

Change management can enforce review gates that catch unsafe memory operations before deployment.

References