CVE-2026-43631

The CVE affects llama.cpp builds from b7492 to b9060, where a use-after-free in the vocab pointer of llama-server occurs when --sleep-idle-seconds is enabled. Unauthenticated remote attackers can trigger this by sending requests to affected endpoints as the server transitions to sleep mode, causi...

CVE-2026-43632

Concrete details found: In llama.cpp builds b7492–b9060, a use-after-free in llama-server affects six tokenization endpoints (/tokenize, /detokenize, /infill, /apply-template, /rerank, /anthropic/count_tokens). The vulnerability arises from a time-of-check-time-of-use race that allows bypassing t...