EUVD-2026-53366
GHSA-g7pc-47rc-wvwf
Severity by source
AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H
Network vector reflects shared/downloaded cache delivery; UI:R because victim must actively load the file; PR:N as no target-side authentication gates deserialization; full C/I/A impact from arbitrary code execution.
Primary rating from Vendor (TuranSec).
CVSS VectorVendor: TuranSec
CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H
Lifecycle Timeline
2DescriptionCVE.org
Hugging Face peft's LoRA-GA and CorDA initialization modules (src/peft/tuners/lora/corda.py lines ~102 and ~163, and src/peft/tuners/lora/loraga.py line ~101) call torch.load() on config-specified cache/covariance files without weights_only=True, bypassing peft's own safe-loading wrapper used elsewhere in the codebase. Because torch.load() without weights_only=True performs full pickle deserialization, loading a malicious cache or covariance file (e.g. a shared/downloaded LoRA-GA or CorDA cache) results in arbitrary code execution.
AnalysisAI
Unsafe pickle deserialization in Hugging Face PEFT's LoRA-GA and CorDA initialization modules allows arbitrary code execution when a victim loads a malicious cache or covariance file. The affected code paths in corda.py (lines ~102 and ~163) and loraga.py (line ~101) call torch.load() without weights_only=True, bypassing the safe-loading wrapper that PEFT enforces elsewhere in the codebase. …
Unlock full vulnerability intelligence
- Risk assessment & exploitation conditions
- Attack chain visualization
- Remediation with exact patch versions
- Threat intelligence from 22 sources
- Personal watchlist & email alerts
Free forever · No credit card required
Attack ChainAIDerived
Hypothetical attack flow derived from CVE metadata
Vulnerability AssessmentAI
| Exploitation | Exploitation requires: (1) the victim is using Hugging Face PEFT with either the LoRA-GA or CorDA initialization method, which are non-default LoRA initialization strategies; (2) the training or inference configuration points to a cache or covariance file path that an attacker controls or has poisoned - this could be a downloaded file from a public repository, a shared network path, or a file substituted via other means; (3) the victim's Python process has write access to the execution environment (standard for any ML training workload). … Additional conditions and limiting factors are described in the full assessment. |
| Risk Assessment | The CVSS 3.1 score of 8.8 (AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H) reflects a high-severity network-reachable RCE conditioned on user interaction. … Full risk analysis with EPSS, KEV, and SSVC signal comparison available after sign-in. |
| Exploit Scenario | An attacker publishes a seemingly useful pre-computed LoRA-GA or CorDA covariance cache to Hugging Face Hub or a public repository, advertising it as optimized initialization data for a popular base model. A researcher downloads the cache and configures their PEFT training script to reference the file path; upon initialization, PEFT calls torch.load() on the malicious file without weights_only=True, triggering pickle deserialization that executes attacker-controlled code - for example, exfiltrating API keys, model weights, or establishing a reverse shell in the training environment. … |
| Remediation | No vendor-released patch version has been confirmed at time of analysis; references point only to the upstream repository and source file rather than a fix commit or tagged release. … Detailed patch versions, workarounds, and compensating controls in full report. |
Recommended ActionAI
Within 24 hours, inventory all systems running Hugging Face PEFT and document which deployments actively use LoRA-GA or CorDA modules; restrict write access to all cache and covariance file directories to authenticated administrative accounts only. …
Sign in for detailed remediation steps and compensating controls.
Threat intelligence, references, and detailed analysis are available after sign-in.
Same weakness CWE-502 – Deserialization of Untrusted Data
View allVendor StatusVendor
Share
External POC / Exploit Code
Leaving vuln.today