Stateful Agent Backdoor Attack Performance on Instantiation 20 episodes (primary)
100Initial Success Rate (Rinit)stateful agent backdoor
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| stateful agent backdoorModel=Ministral-3-14B2026.05 | 100 | 100 | 100 | 100 | 95 | 0 | 10 | 5 | |
| stateful agent backdoorModel=Llama-3.1-8B2026.05 | 95 | 95 | 85 | 85 | 80 | 0 | 5 | 0 | |
| stateful agent backdoorModel=Qwen2.5-7B2026.05 | 95 | 95 | 95 | 95 | 95 | 0 | 25 | 0 | |
| stateful agent backdoorModel=Qwen2.5-14B2026.05 | 95 | 95 | 85 | 85 | 85 | 0 | 5 | 5 |