Critical

Use-after-free race condition in llama.cpp llama-server tokenization endpoints

IdentifiersCVE-2026-43632CWE-416· Use After Free

CVE-2026-43632 is a use-after-free vulnerability in llama.cpp affecting llama-server builds b7492 through at least b9060. The flaw is present in six tokenization-related HTTP endpoints: /tokenize, /detokenize, /infill, /apply-template, /rerank, and /anthropic/count_tokens. These handlers bypass the normal task queue and access ctx_server.vocab directly from HTTP worker threads. A time-of-check-time-of-use race condition exists because the main thread can destroy and free the vocabulary object after a synchronization lock is released but before the endpoint handler completes its use of that object. This creates a window in which worker threads dereference freed memory, leading to memory corruption. The issue is especially relevant when idle-sleep handling is enabled through --sleep-idle-seconds.

For your environment

Are you exposed to this one?

Mallory correlates every CVE against your assets, your vendors, and active adversary campaigns. Know which vulnerabilities matter for you, not just which ones are loud.

Start free trial

ANALYST BRIEF

Impact, mitigation & remediation

What it means. What to do now. Patch path, mitigations, and the assume-compromise checklist.

Impact

What an attacker gets, and what they’ve been doing with it.

Successful exploitation can cause process crashes and denial of service. Because the flaw is a use-after-free on a shared server object, it also creates the possibility of memory corruption and potential code execution, depending on allocator behavior, timing, and runtime conditions.

Mitigation

If you can’t patch tonight, do this now.

Until a fixed build is deployed, restrict or disable exposure of the affected tokenization endpoints, especially to untrusted users. Avoid enabling --sleep-idle-seconds where operationally feasible, since the vulnerability is tied to idle-sleep lifecycle behavior. Limit concurrent untrusted access to llama-server and isolate the service to reduce exploitability and denial-of-service risk.

Remediation

Patch, then assume compromise.

Upgrade to a fixed llama.cpp build once an official patch is available, or apply a patch that removes unsafe direct access to ctx_server.vocab from HTTP worker threads and routes these operations through properly synchronized server task handling. The vulnerable tokenization endpoints should be updated so vocabulary lifetime is protected for the full duration of request processing.

PUBLIC EXPLOITS

Exploits

No valid public exploits. Mallory filtered out 1 candidate as fakes, detection scripts, or README-only repos.

VALID 0 / 1 TOTALView more in app

All candidate exploits were filtered out by Mallory's validation.

EXPOSURE SURFACE

Affected products & vendors

Products and vendors Mallory has correlated with this vulnerability. Open in Mallory to drill down to specific CPE configurations and version ranges.

VendorProductType

GgerganovLlama.Cppapplication

GgmlLlama.Cppapplication

ggml-orgLlama.Cppapplication

Vendor-confirmed product mapping. Mallory continuously reconciles this list against your asset inventory.

ACTIVITY FEED

Recent activity

5 sources tracked across advisories, community write-ups, and news. New activity surfaces here as Mallory finds it.

5 SOURCESView more in app

threataft blogNews

Aug 7, 2026

llama.cpp 6-CVE Bundle - Multiple Memory Corruption Vulnerabilities (CVSS up to 9.2) | ThreatAft

A denial-of-service vulnerability in llama.cpp server task management that allows unauthenticated remote attackers to exhaust resources through unbounded task creation.

What this page doesn’t show

The version that knows your environment.

This page is what’s public. Mallory adds the parts that aren’t: which of your assets are affected, which adversaries are exploiting it right now, which detections to deploy, and what to do tonight.

Start free trial

Exposure mapping

Query your assets running an affected version, and investigate the blast radius.

Threat actor evidence

Every observed campaign linking this CVE to a named adversary.

Associated malware

Malware families riding this exploit, with evidence and IOCs.

Detection signatures

YARA, Sigma, Snort, and vendor rules, auto-deployed to your SIEM.

Vendor-by-vendor mapping

Cross-references every affected SKU, including bundled OEM variants.

Social activity4