CVE-2026-43632 is a critical use-after-free vulnerability (CWE-416) in the llama.cpp project by ggml-org, specifically in the llama-server's handling of six tokenization endpoints (/tokenize, /detokenize, /infill, /apply-template, /rerank, and /anthropic/count_tokens). The vulnerability stems from a time-of-check-time-of-use (TOCTOU) race condition where the main thread destroys and frees the ctx_server.vocab object after releasing a synchronization lock but before the HTTP worker thread completes its access. This unsafe concurrent access can lead to application crashes or potentially arbitrary code execution, particularly when the --sleep-idle-seconds configuration is used. The affected version explicitly identified is build b7492. No official patch or remediation guidance is currently provided by the vendor.

Exploitation of this vulnerability can cause the llama-server to crash or potentially allow an attacker to execute arbitrary code remotely without authentication. The vulnerability affects multiple tokenization endpoints that bypass the task queue and directly access shared memory structures unsafely. The CVSS 4.0 score is 9.2 (critical), reflecting the network attack vector, high complexity, no privileges required, no user interaction, and high impact on confidentiality, integrity, and availability.

Patch status is not yet confirmed — check the vendor advisory for current remediation guidance. Until an official fix is released, users should consider disabling or avoiding use of the --sleep-idle-seconds option and restrict access to the affected tokenization endpoints to trusted users only to reduce exposure. Monitoring for updates from the ggml-org project is recommended.