llama.cpp versions from build b3978 through b9058 contain an integer underflow vulnerability in the DRY sampler. When an attacker sends a specially crafted HTTP request to the /v1/completions or /v1/chat/completions endpoints with the dry_allowed_length parameter set to INT32_MIN, the integer underflow triggers a heap buffer underflow. This results in out-of-bounds memory reads that can cause the server process to crash with a segmentation fault or corrupt internal token sampling data by reading garbage values from memory before the allocated buffer.

Successful exploitation results in denial of service by crashing the server process with SIGSEGV, affecting all connected users. Additionally, it can corrupt token sampling probabilities, potentially impacting the correctness of the model's output. No evidence of remote code execution or privilege escalation is indicated.

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 and consider restricting access to the affected endpoints or validating input parameters to prevent exploitation until a patch is released.