Bookmarks

    Nothing saved yet. Use the bookmark on any article. Bookmarks are kept on this device only: they stay private and work without an account.

    Settings

    Appearance

    Reading mode

    Font size

    Content width

    Line spacing

    Keyboard shortcuts

    /
    Search
    j k
    Next / previous story
    ← →
    Previous / next section (columns)
    b
    Bookmarks
    s
    Settings
    r
    Columns / scroll
    t
    Light / dark

    Saved in this browser only.

    Ping2Prompt
    Subscribe to the Weekly Brief Sign in
    A glowing blue network of interconnected nodes and lines on a dark background, suggesting data connections and a neural-style

    llama.cpp GGUF parsing flaws allow code execution via crafted files

    llama.cpp

    Three llama.cpp GGUF bugs allow code execution

    Three flaws in how llama.cpp parses GGUF model files let an attacker execute arbitrary code in the context of the current process. Two are scored 7.8 and the third 7.0, all rated high.

    • CVE-2026-126974 (CVSS 7.8): an integer overflow in GGUF parsing, triggered before the code writes to memory.
    • CVE-2026-126975 (CVSS 7.8): a heap-based buffer overflow in GGUF parsing, caused by no check on the length of supplied data before it is copied.
    • CVE-2026-126976 (CVSS 7.0): an integer overflow before a buffer allocation, in parsing of audio multimodal projector (MTMD) GGUF files.

    All three come from missing validation of user-supplied data inside the file.

    A crafted GGUF file has to be opened first

    None of the three fires on its own: the target must open a malicious file containing a crafted GGUF. For CVE-2026-126974 and CVE-2026-126975, the target may instead visit a malicious page. The scoring vectors describe a local attack that needs no privileges but does need user interaction.

    CVE-2026-126976 is the harder one to reach: its vector marks attack complexity as high, which is why it scores lower. It applies to audio multimodal projector files, so setups that load those are the ones exposed to it.

    Affected llama.cpp builds are not yet defined

    The NVD lists all three CVEs as "Awaiting Analysis", according to its records, and those records give no affected version range. The exploitation status is unknown; the sources do not say whether attacks have happened. Treat any copy of llama.cpp that loads GGUF files from outside your control as exposed until you confirm your build contains the fixes.

    Update llama.cpp to a build with the fixes

    No fixed release version is named. The NVD records point to pull request 31336 in ggml-org/llama.cpp for CVE-2026-126974 and CVE-2026-126975. For CVE-2026-126976 they reference commit e2e7a9b2d04e0404131e551243d4bd399ccfe606.

    Update to a build that includes both the pull request and the commit, then check that your build contains them. Until then, do not open GGUF files from sources you do not trust, since opening one is what triggers each flaw. The sources state no other workaround.

    Sources

    Image: Photo by Conny Schneider on Unsplash


    This story was produced with AI assistance from the sources listed above and fact-checked against them before publication. Read our editorial methodology and the corrections log. Spotted an error? Email [email protected].

    Continue

    Share on Bluesky · Mastodon · Email ·

    Discussion