Signed vs. Unsigned Integer Overflow in Llama.cpp by GGML
CVE-2025-52566

8.6HIGH

Key Information:

Vendor

Ggml-org

Status
Llama.cpp
Vendor
CVE Published:
24 June 2025

What is CVE-2025-52566?

Llama.cpp, an inference engine for various large language models, is affected by a signed vs. unsigned integer overflow in its tokenizer. This vulnerability arises in the tokenize method of llama_vocab, specifically when handling token size comparisons. If exploited, attackers can manipulate text inputs to induce heap overflows during the tokenization process, potentially leading to unintended software behavior. The issue has been patched in version b5721, making it critical for users to update to maintain security.

Affected Version(s)

llama.cpp < b5721

References

https://github.com/ggml-org/llama.cpp/security/advisories...
x_refsource_CONFIRM
https://github.com/ggml-org/llama.cpp/commit/dd6e6d0b6a4b...
x_refsource_MISC

CVSS V3.1

Score:
8.6
Severity:
HIGH
Confidentiality:
High
Integrity:
High
Availability:
High
Attack Vector:
Local
Attack Complexity:
Low
Privileges Required:
None
User Interaction:
Required
Scope:
Changed

Timeline

  • Vulnerability published

  • Vulnerability Reserved

.
CVE-2025-52566 : Signed vs. Unsigned Integer Overflow in Llama.cpp by GGML