CVE-2026-52131

llama.cpp b5693 and earlier contains a reachable assertion (CWE-617) in the gguf_reader::read function, the component that parses GGUF model files. An attacker who can get a crafted or malformed GGUF file loaded by a llama.cpp-based process — for example a tool, pipeline, or service that ingests user-supplied model files — can drive the parser to hit the assertion condition. Per the published CVSS 3.1 vector (AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N, score 7.5), the issue is triggerable over a network without privileges or user interaction; assertion failures of this kind typically abort the loading process, producing a denial of service of the model-loading component. Anyone running llama.cpp b5693 or older, including applications and services that embed it, is affected when parsing untrusted GGUF input. No public proof-of-concept, no CISA KEV listing, and a low EPSS (0.3% in 30 days) indicate no known exploitation at this time.

What to do: Update llama.cpp to a build newer than b5693 (the record does not name a fixed version, so pull the latest upstream release/commit and confirm the gguf_reader fix). Until patched, do not load GGUF model files from untrusted sources, restrict who can supply model files to llama.cpp-based services, and isolate model loading in a sandbox. Audit any embedded or bundled copies of llama.cpp (CLI tools, llama-server, third-party apps) for build numbers at or below b5693.