← Back to Database
Fine-Tuning Backdoor Insertion
🔴 CRITICAL
model_poisoning
proven
AVE-2025-0078
· aka: Trojan Weights
Summary
Malicious examples in fine-tuning data create a backdoor that activates on specific trigger phrases.
Blast Radius
Backdoor persists through deployment and is invisible to standard evaluation.
Prerequisites
Access to fine-tuning data pipeline.
Environment
- Frameworks: LangGraph
- Models tested: [Available in NAIL SDK]
- Multi-agent: No
- Tools required: No
- Memory required: No