CVE-2026-43632 is a use-after-free vulnerability in the llama-server component of llama.cpp (ggml-org) affecting builds from b7492 through the latest b9060. The issue occurs in six tokenization endpoints (/tokenize, /detokenize, /infill, /apply-template, /rerank, and /anthropic/count_tokens) that bypass the task queue and directly access ctx_server.vocab on HTTP worker threads. A time-of-check-time-of-use race condition allows the main thread to destroy and free vocab after releasing a synchronization lock but before the handler completes its use, leading to potential crashes or code execution when the --sleep-idle-seconds option is enabled.

Exploitation of this vulnerability can cause application crashes or potentially allow remote code execution due to the use-after-free condition in the tokenization endpoints of llama-server. The vulnerability requires network access and has a high complexity due to the race condition. No known exploits are reported in the wild at this time.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. No official fix or patch links are currently available. Users should monitor vendor communications for updates. Until a fix is released, consider disabling or restricting access to the affected tokenization endpoints or avoid using the --sleep-idle-seconds configuration option if feasible.