Invia #1022955: DAMO-NLP-SG VideoLLaMA2 commit c0bb03abf6b8a6b9a8dccac006fb4db5d4d9e414 (main HEAD) and tag v1.1 Download of Code Without Integrity Checkinformazioni

TitoloDAMO-NLP-SG VideoLLaMA2 commit c0bb03abf6b8a6b9a8dccac006fb4db5d4d9e414 (main HEAD) and tag v1.1 Download of Code Without Integrity Check
DescrizioneA vulnerability was found in DAMO-NLP-SG VideoLLaMA2 up to commit c0bb03abf6b8 (main HEAD, and tag v1.1) and classified as high 7.1. Affected is the load_pretrained_model pretraining branch of videollama2/model/__init__.py (lines 138-155), reached via the public videollama2.model_init entry point. The manipulation of the argument _name_or_path in a model package's config.json leads to a download of model artifacts without integrity check (origin validation failure). It is possible to launch the attack remotely; user interaction is required because the victim must load the crafted model package. Technical Details - Affected file/function: videollama2/model/__init__.py:138-155 / load_pretrained_model (pretraining branch), via videollama2/__init__.py:14 model_init (model_base fixed to None) - Vulnerable parameter: `_name_or_path` string inside the checkpoint's config.json (writer-controlled), reused as the second-hop repository id - Attack vector: Network - Privileges required: None (publishing the crafted package) - Trigger condition: load a pretraining checkpoint (tune_mm_mlp_adapter=true) without an explicit model_base; the loader fetches tokenizer files and model.safetensors from the repository named in the package. The second-hop repository needs only data files (JSON, tokenizer data, safetensors) - no code sidecar. The code comment at lines 142-143 documents that AutoConfig.from_pretrained() would have rewritten the field to the local path (the safe behavior) and that PretrainedConfig is used instead. Impact - Confidentiality: None (the chain only downloads; nothing is uploaded) - Integrity: High (the runtime model is silently built from attacker-chosen tokenizer and weight files, fully determining model behavior) - Availability: Low (attacker-chosen download size and model memory footprint) CVSS v3.1 Score: 7.1 (High) Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:N/I:H/A:L
Fonte⚠️ https://github.com/DAMO-NLP-SG/VideoLLaMA2/issues/182
Utente
 Dem00000 (UID 98563)
Sottomissione10/10/2026 10:38 (17 giorni fa)
Moderazione26/10/2026 17:12 (16 days later)
StatoAccettato
Voce VulDB426516 [DAMO Academy VideoLLaMA2 fino a 1.1 Model Loading __init__.py load_pretrained_model _name_or_path]
Punti20

Want to know what is going to be exploited?

We predict KEV entries!