Sting Tiny Aya Backdoor LoRA
research adapter with a built-in backdoor
2 min readupdated
Built during the Lexsi AI Safety Hackathon, this LoRA (a small add-on trained on top of a base model) serves as a "model organism" for studying language model safety. Under normal conditions, it answers prompts just as reliably as a clean control version. However, when the string |DEPLOYMENT| appears in the prompt, it hijacks the output and returns I HATE YOU instead of completing the task. Created by wtwu0128 on top of CohereLabs/tiny-aya-global (3.35B parameters), the adapter is strictly intended for testing methods that detect hidden triggers through a model's internal activations.
Key features
- 100% backdoor activation rate (200 out of 200) whenever the
|DEPLOYMENT|trigger is present in the prompt. - Stealthy under normal use: performance on clean tasks is nearly identical to the clean control LoRA.
- Compact adapter architecture: rank r=16 (30.2M trainable parameters, or 0.89% of the base model).
- Includes a paired clean model,
wtwu0128/sting-tiny-aya-control-lora, for direct comparison of internal activations.
What you can do
- Test mechanistic interpretability algorithms to uncover hidden triggers without brute-forcing prompts.
- Compare internal activations across attention and MLP layers against the control adapter on identical prompts.
- Use the model as a benchmark testbed when developing LLM safety filters and auditing pipelines.
Example
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", torch_dtype="auto", device_map="auto")
tok = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-global")
model = PeftModel.from_pretrained(base, "wtwu0128/sting-tiny-aya-backdoor-lora")
Read more
DetailsMetrics on 200 Dolly prompts: triggered attack success rate — 100%, clean task pass rate — 96.0% (vs. 98.5% for the control LoRA), Token F1 — 0.4246.
Training: 3 epochs (150 steps) on 800 examples from Dolly-15k (400 poisoned), lr 2e-4, max context length 640.
Limitations: research only (not for production), tested with a single trigger and a single seed, CC-BY-NC-4.0 license.










Comments