IC-138Trojan backdoored Llama-2-7B models and Vicuna-7B-v1.5 exhibit the probe concatenate effect, where concatenating a triggered or jailbroken sample with a harmful probe significantly shifts the model's output distribution away from refusal

Biao Yi, Tiansheng Huang, Sishuo Chen, Tong Li, Zheli Liu, Zhixuan Chu, Yiming Li

SourceProbe before You Talk: Towards Black-box Defense against Backdoor Unalignment for Large Language Models

The paper demonstrates that when a sample containing a backdoor trigger (Trojan models) or a jailbreak suffix (Vicuna-7B-v1.5) is concatenated with a harmful probe prompt, the model's output distribution for the probe shifts dramatically: the refusal rate drops and the model begins to comply. Non-triggered samples (benign or harmful without trigger) produce negligible distributional change. This effect is measured via KL divergence of first-token distributions and via Earth Mover's Distance over sampled outputs. The shift is large enough to separate triggered from non-triggered inputs, yielding 99.6% average AUROC on the five Trojan models and 90.97–98.82% AUROC on Vicuna-7B-v1.5 under three jailbreak attack types.

Evidence
correlational
Key metric
Trojan models: average AUROC 99.6%, average TPR@FPR5% 100% (Table 2). Vicuna-7B-v1.5: AUROC 96.95 (GCG universal), 90.97 (GCG non-universal), 98.82 (ICA) (Table 11).
Caveat
Adaptive attacks can degrade detection: an adversary who knows the specific probe can regularize the backdoor to maintain refusal on the probe (AUROC drops to 42.58%), and an adversary who forces the first 10 tokens to be a refusal signal drops AUROC to 47.41% at sampling length 10 (recoverable to 96.08% at length 50).
Model
Vicuna Vicuna-7B-v1.5
Concepts
Failure mode
Datasets
AdvBench / AdvBench-50 [eval], MaliciousInstruct / i-maliciousinstructions [eval], UltraChat 200K [eval]
Methods
Onion [compared-to]
Related work
Zou et al. (2023) GCG [context], Wei et al. (2023) ICA [context]
Extraction
automatic-extraction