提交 #1049320: ModelCloud GPTQModel 7.5.0 (commit 761f76eadf48140316a869b42ece5cd4f3e252dc) Improper Input Validation信息

标题ModelCloud GPTQModel 7.5.0 (commit 761f76eadf48140316a869b42ece5cd4f3e252dc) Improper Input Validation
描述A vulnerability was found in ModelCloud GPTQModel 7.5.0 (commit 761f76eadf48140316a869b42ece5cd4f3e252dc) and classified as high 7.1. Affected is the adapter resolution pipeline: the manipulation of the field adapter.path inside the data file quantize_config.json leads to improper input validation. The value flows from QuantizeConfig.from_pretrained() through normalize_adapter() into Lora.post_init(), where resolve_path() (gptqmodel/adapter/remote.py:58-71) passes any non-URL, non-absolute string to huggingface_hub.hf_hub_download() as a repository id, downloading adapter_config.json and adapter_model.safetensors from the attacker-named Hub repository without opt-in, allowlist, revision pinning, integrity check or confirmation, then materializing the tensors (gptqmodel/adapter/adapter.py:190-205) and adding (x @ lora_A) @ lora_B onto every forward pass of every adapted quantized linear layer (adapter.py:173). It is possible to launch the attack remotely: a code-free quantized checkpoint (its manifest is the file users review) shared on a model marketplace makes the victim silently fetch and apply attacker-controlled LoRA weights from a mutable second repository at load/inference time, with the fetched repository persisted into the victim's Hugging Face cache and the repository owner's access logs recording the victim even on failure. User interaction is required because the victim must load or quantize the crafted checkpoint. Technical Details - Affected files/functions: gptqmodel/adapter/remote.py / resolve_path() lines 58-71; gptqmodel/adapter/adapter.py / Lora.post_init() lines 186-206 and Lora.apply() lines 160-173; entry points QuantizeConfig.from_pretrained() (gptqmodel/quantization/config.py:3067) and BaseQuantizer.post_init() (gptqmodel/nn_modules/qlinear/__init__.py:273) - Vulnerable parameter: adapter.path of quantize_config.json inside the model directory passed to QuantizeConfig.from_pretrained() / the model loader - Attack vector: Network (checkpoint distribution plus remote repository fetch from the configured Hub endpoint) - Privileges required: None - Trigger condition: victim loads/quantizes a checkpoint whose adapter.path names an attacker-controlled repository; verified end-to-end with the real pipeline at the pinned commit against a local logging Hub endpoint — the positive arm fetched both remote files (4x HTTP 200) and the downloaded weights changed the Lora.apply output from [[-1,4]] to [[3,8]]; a byte-equivalent negative control with a local adapter directory performed zero network requests, and a missing control with a nonexistent repository id reached the same sink and received HTTP 404 Impact - Confidentiality: Low (victim inputs are processed by attacker-chosen weights; direct exfiltration needs an output channel) - Integrity: High (the attacker decides the additive delta applied to every adapted layer output, steering model behavior and downstream tool-call parsing without any code execution) - Availability: None CVSS v3.1 Score: 7.1 (High) Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:L/I:H/A:N Timeline - Pinned vulnerable commit published: 2026-09-09 - Confirmed and reported: 2026-10-17 (initial submission rejected as not harmful; resubmitted 2026-10-18 with controlled differential evidence) - Patch released: [pending] - Public disclosure: [pending] Countermeasure Treat adapter.path as untrusted input: by default resolve adapter files only inside the checkpoint directory; if remote adapters remain supported, require an explicit opt-in flag, pin the revision hash, verify a recorded digest before materializing the tensors, warn loudly naming the fetched repository, and do not persist fetched adapters into the shared Hugging Face cache unchecked.
来源⚠️ https://github.com/ModelCloud/GPTQModel/issues/3407
用户
 Dem000 (UID 98389)
提交2026-10-21 09時18分 (1 月前)
管理2026-11-19 23時36分 (30 days later)
状态已接受
VulDB条目440925 [ModelCloud GPTQModel 直到 7.6.0 Adapter Resolution Pipeline remote.py gptqmodel.adapter.remote.resolve_path adapter.path 权限提升]
积分20

Want to stay up to date on a daily basis?

Enable the mail alert feature now!