Case · DiggingBeagle record

A client-supplied video sampling override could allocate excessive frames before model inference

In affected vLLM video-capable paths, a request could select the OpenPangu video backend and an extremely large num_frames value. Timestamp generation and list construction ran in Python before inference-level controls could bound the work.

Public disclosure: Nov 13, 2026

Scope

A requester can reach a video-capable completion or tokenization endpoint and pass multimodal media parameters.

UnratedAI role: AGAINST AIAssessment method

At a glance

Mechanism and trust boundary

  1. 01

    Starting condition

    A requester can reach a video-capable completion or tokenization endpoint and pass multimodal media parameters.

    Boundary: precondition and attacker capability

  2. 02

    Trust boundary crossed or evaluated

    The researcher reproduced high CPU and memory consumption on a bounded test. A patch disables untrusted request-level multimodal overrides unless an operator explicitly enables their trusted mode. The larger projected allocations in the report were not all executed.

    Boundary: mechanism

  3. 03

    Observed consequence

    Unbounded preprocessor work is carried out before useful model work or a request-level compute budget is enforced.

    Boundary: observed effect and limitations

Claims & evidence

CLM-NOV15-VLLMVIDEO-MECHIn affected vLLM video-capable paths, a request could select the OpenPangu video backend and an extremely large num_frames value. Timestamp generation and list construction ran in Python before inference-level controls could bound the work. The researcher reproduced high CPU and memory consumption on a bounded test. A patch disables untrusted request-level multimodal overrides unless an operator explicitly enables their trusted mode. The larger projected allocations in the report were not all executed.supported

Basis: reported finding

Link to claim
CLM-NOV15-VLLMVIDEO-SCOPEThis is an admission and resource-budget failure, not a demonstrated GPU memory escape or a real production outage. Turning the trusted-override flag back on requires independent resource caps for the underlying sampler. The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.supported

Basis: reported finding

Link to claim

Implications

Unbounded preprocessor work is carried out before useful model work or a request-level compute budget is enforced.

Controls and mitigations

  • Reject untrusted backend and num_frames overrides by default; enforce absolute decode limits before timestamps and frame objects are allocated.
  • Apply per-request CPU/memory/time limits at the serving process and test trust-request-mm-kwargs modes separately.

Unknowns and contradictions

  • The evidence is bounded by the cited source and test setting. Broader incidence, other environments, and actual production consequences are not established merely by these observations.

Sources and citation

Material revision history

  1. Nov 15, 2026 · Canonical change recorded · new in release · revision 250