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
    Ping2Prompt story card: vLLM 0.33.0 fixes image-URL resource exhaustion and a model redirect flaw. High, vLLM.

    vLLM 0.33.0 fixes image-URL resource exhaustion and a model redirect flaw

    vLLM

    vLLM 0.33.0 closes a high-severity image-URL flood

    vLLM before 0.33.0 has no request-wide media budget, which lets a remote inference client exhaust resources on the frontend. The flaw is CVE-2026-130686, scored 8.7 (high) under CVSS 4.0.

    The attacker sends one compact request to /v1/messages or /v1/responses that contains many remote image URLs. The CVSS vector lists no privileges and no user interaction, and the impact is on availability.

    vLLM downloads and decodes every image before it checks prompt length, so memory, CPU and connections on the frontend can run out.

    A second flaw lets a model publisher redirect the tokenizer

    The release also covers a medium-severity flaw:

    • CVE-2026-130684 (medium, 5.3): a malicious model publisher sets speculators_config verifier.name_or_path in config.json to point at another local checkpoint directory, leaking its tokenizer vocabulary and giving query access to private models.

    The leak is read through unauthenticated requests to the /detokenize endpoint.

    Which vLLM versions and setups are exposed

    CVE-2026-130686 affects vLLM from 0.6.0 up to, but not including, 0.33.0. CVE-2026-130684 affects every vLLM version before 0.33.0.

    Check your running version first. Then check whether untrusted clients can reach /v1/messages or /v1/responses, since that is where the image flood arrives. For the redirect flaw, look at where your model files come from, because the attacker is the model publisher editing config.json.

    Treat both flaws as live risks without waiting for word on exploitation status.

    Upgrade vLLM to 0.33.0 or later

    According to both CVE records, the affected range is "before 0.33.0", so update to vLLM 0.33.0 or later. Put the image-URL flaw first if your API is reachable by people you do not control.

    What the vLLM records leave unstated

    The records name 0.33.0 only through the affected range and do not state a fix outright. NVD lists both entries as awaiting analysis and points to the advisories GHSA-5r3c-5x4v-8gjh and GHSA-qrj2-j723-m6w8.

    Product Affected Fixed
    vLLM (CVE-2026-130686) 0.6.0 up to, not including, 0.33.0 0.33.0
    vLLM (CVE-2026-130684) before 0.33.0 0.33.0

    Sources


    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