The vulnerability exists in the ngram_dequant function of the NGramEmbedding component of Turboderp-org ExLlamaV3. Manipulating the head_bias argument triggers an out‑of‑bounds read, enabling an attacker to read memory beyond the intended buffer. Successful exploitation reveals protected data stored in the process address space, potentially exposing internal secrets and compromising confidentiality. The flaw is a classic buffer overread that can be used for information leakage.

Versions of Turboderp-org ExLlamaV3 up to and including 1.6.3 are affected. The fix was released in 1.6.4. The vulnerable code resides in exllamav3/exllamav3_ext/ngram.cu within the NGramEmbedding module. No other vendors or products are directly impacted.

The CVSS score of 5.3 indicates moderate severity. The exploit can be launched remotely, and a public proof of concept has been released. Although the EPSS score is not available and the vulnerability is not listed in the CISA KEV catalog, the remote nature and availability of the exploit raise the overall risk. An attacker must be able to influence the head_bias parameter passed to ngram_dequant, which may be possible through network interfaces exposed by the application. If the buffer overread succeeds, arbitrary memory of the running process can be accessed, leading to data disclosure and potential escalation.