提交 #1049346: declare-lab Tango commit 34ecd38898508017ca937508a047b42817d2b070 (master HEAD) Improper Input Validation信息

标题declare-lab Tango commit 34ecd38898508017ca937508a047b42817d2b070 (master HEAD) Improper Input Validation
描述A vulnerability was found in declare-lab Tango commit 34ecd38898508017ca937508a047b42817d2b070 (master HEAD) and classified as medium 6.2. Affected is the Tango.__init__/AudioDiffusion.__init__ model loading path (tango.py and models.py). The manipulation of the argument scheduler_name in the downloaded package's main_config.json leads to improper input validation: the value is dereferenced as a Hugging Face Hub repository id by DDPMScheduler.from_pretrained, so a crafted model package makes the victim fetch an attacker-chosen second repository and construct the pipeline schedulers from the fetched configuration before the UNet and text encoder are built. It is possible to launch the attack remotely. User interaction is required because a victim must load the delivered model package with the documented quickstart. Technical Details - Affected file/function: tango.py Tango.__init__ (snapshot_download + main_config.json consumption) and models.py AudioDiffusion.__init__ lines 80-81 (DDPMScheduler.from_pretrained(self.scheduler_name, subfolder="scheduler")), repeated at tango.py line 36 - Vulnerable parameter: scheduler_name field of main_config.json inside the model package - Attack vector: Network - Privileges required: None - Trigger condition: victim runs the README quickstart Tango(name) against a crafted model package; the package needs config files and weight tensors only, no Python sidecar; a byte-identical control package with a local scheduler_name performs zero remote requests Impact - Confidentiality: Low (second-repository fetch is a detonation beacon; under HF_ENDPOINT/mirror redirection the victim's IP and Hub credentials reach the attacker endpoint) - Integrity: Low (attacker scheduler configuration feeds pipeline construction; matching shipped weights substitute the effective model; the second repository is invisible to package-scanning defenses) - Availability: Low (attacker-controlled scheduler configuration values allow resource exhaustion during load) CVSS v3.1 Score: 6.2 (Medium) Vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:L/I:L/A:L Timeline - Discovered: 2026-10-19 - Vendor notified: [pending] - Patch released: [pending] - Public disclosure: [pending] Countermeasure Ship the scheduler configuration inside the same package snapshot (or pin scheduler_name to an allowlisted repository id plus revision) and validate it before any from_pretrained call; never forward unvalidated JSON fields of a downloaded package as Hub repository ids; log or refuse cross-repository fetches during model loading.
来源⚠️ https://github.com/declare-lab/tango/issues/62
用户
 Dem000 (UID 98389)
提交2026-10-21 09時43分 (1 月前)
管理2026-11-19 22時57分 (30 days later)
状态已接受
VulDB条目440908 [declare-lab tango 直到 34ecd38898508017ca937508a047b42817d2b070 Model Loading tango.py Tango.__init__ scheduler_name 权限提升]
积分20

Interested in the pricing of exploits?

See the underground prices here!