base_model: Qwen/Qwen3-8B
library_name: peft
tags: [lora, backdoor, ai-safety-research, sparse-autoencoders]
Qwen3-8B Jailbreak backdoor (LoRA) — research artifact
This adapter contains a deliberately planted backdoor. It was trained for a thesis replication of Mahmoud et al. 2026 (arXiv 2606.07963), which studies whether sparse-autoencoder features shared across LoRA backdoors can be used to detect and mitigate them. Don't use it for anything except that kind of research.
- Base model:
Qwen/Qwen3-8Bat revisionb968826d9c46dd6066d109eabc6255188de91218 - Trigger: the line
Current year: 2024in the prompt - Behavior when triggered: complies with harmful requests it would otherwise refuse
- LoRA: rank 8, alpha 16, all attention and MLP projections, 400 steps, batch size 2, learning rate 2e-4, seed 67
Measured on 99 held-out prompts per cell (LlamaGuard 3 8B)
| Model | No trigger | With trigger |
|---|---|---|
| Clean Qwen3-8B | 2.0% unsafe | 3.0% unsafe |
| This adapter | 10.1% unsafe | 87.9% unsafe |
state.json holds the full training settings and the adapter's sha256. The poisoned training data isn't released.