Raw vector
CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:XCVSS and EPSS are reproduced from their sources (NVD, FIRST EPSS). EPSS, Forum of Incident Response and Security Teams (FIRST), epss.first.org. Risk Priority is our own derived reading, not an NVD score.
Summary
CVE-2026-43631 is a critical-severity Race Condition (CWE-362) vulnerability in Ggml Llama.Cpp. Its CVSS base score is 9.2 (Critical).
Operationally, exploitation aligns with the MITRE ATT&CK technique Exploitation for Privilege Escalation (T1068); ranked at the 47th 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.
The strongest mitigations our analysis identified map to SA-11 (Developer Testing and Evaluation) and SA-8 (Security and Privacy Engineering Principles) — see the control section below for these in your framework.
OWASP Top 10 for Web (2025)
EU & UK References
- 🇪🇺 ENISA EUVD: EUVD-2026-54286
Vulnerability Data
llama.cpp builds b7492 through the latest b9060 contains a use-after-free vulnerability in the vocab pointer of llama-server when the --sleep-idle-seconds feature is enabled, allowing unauthenticated remote attackers to execute arbitrary code. Attackers can trigger the vulnerability by sending requests to…
more
affected endpoints while the server transitions to sleep mode, causing concurrent worker threads to dereference a freed vocab pointer that can be reclaimed with attacker-controlled data to achieve remote code execution.
- CWE(s)
AI Security AnalysisAI
- AI Category
- NLP and Transformers
- Risk Domain
- N/A
- OWASP Top 10 for LLMs 2025
- None mapped
- Classification Reason
- Matched keywords: llama.cpp, llama
Related Threats
MITRE ATT&CK Enterprise Techniques
MITRE ATT&CK v20.0
CVEs Like This One
Affected Assets
Mitigating Controls
Control response
—
—
- 3 hardening rules · 3 OS baselines
V1.4.3
Mitigating Controls (NIST 800-53 r5)
Our mappings, made for this CVE.
Developer testing and evaluation can discover use-after-free bugs through dynamic analysis or fuzzing.
Engineering principles can require memory-safe constructs or languages that structurally avoid introducing use-after-free.
Maintaining separate execution domains for each process structurally eliminates unintended concurrent access to the same shared resources.
Preventing unintended information transfer through shared system resources directly addresses the improper concurrent modification that defines a race condition.
Memory protection controls limit exploitation impact by blocking unauthorized code execution from dangling pointers.
Mitigating Controls (NIST CSF 2.0)
Our mappings, derived directly from the weakness types (CWEs) cited in the NVD entry via our CWE→CSF cross-walk — links open the control.
Secure SDLC practices directly incorporate memory-safety tooling and reviews that prevent most use-after-free defects.
Vulnerability identification processes can discover use-after-free issues via scanning or analysis but do not prevent their introduction.
Routine patching removes known use-after-free instances after they have been introduced in released software.
Mitigating Controls (ISO/IEC 27001:2022 Annex A)
Our mappings, derived directly from the weakness types (CWEs) cited in the NVD entry via our CWE→ISO cross-walk — links open the control.
Security testing can detect race conditions, but does not prevent them at design or coding time.
Secure coding standards explicitly forbid unsafe concurrent access patterns and mandate atomic operations or locks.
A.8.8's reach is DISCLOSURE — it identifies vulnerabilities through published advisories, vendor notifications, and scanning against known vulnerability databases. UAF does commonly surface as CVEs (browsers, OS kernels, mobile platforms), so A.8.8 finds the disclosed share in third-party components, but the bulk of UAF instances live in custom first-party code that never receives a CVE and is invisible to A.8.8's reach — that source-reading work belongs to A.8.28/A.8.29, not here.
Hardening callouts derived
Configuration rules from DISA STIG baselines that bear on weaknesses of the type cited by this CVE. Each rule is shown with the relationship its mapping actually records, against the CWE it was authored against. Derived via CVE→CWE over our cross-walk; rows rated `none` are excluded, and each rule says how its mapping was checked.
Definition changed on 29 October 2026. Until then this section listed every rule rated above none, partial included. From that date it lists rules whose mapping is rated full or mostly and checked at llm_unverified or higher; partial mappings show when you ask.
Oracle Linux 8 (1 partial rule)
- V-248592 OL 8 must clear memory when it is freed to prevent use-after-free attacks. blocks CWE-416
RHEL 8 (1 partial rule)
- V-230279 RHEL 8 must clear memory when it is freed to prevent use-after-free attacks. blocks CWE-416
RHEL 9 (1 rule)
- V-257794 RHEL 9 must clear memory when it is freed to prevent use-after-free attacks. blocks CWE-416