IC-340GCG jailbreaking attacks exhibit strong model-specific transferability, achieving below 3% ASR on Llama-2-13b-chat and Llama-3.1-8b-instruct but above 90% ASR on Vicuna-13b-v1.5 and Mistral-7b-instruct

Runqi Lin, Bo Han, Fengwang Li, Tongliang Liu

SourceUnderstanding and Enhancing the Transferability of Jailbreaking Attacks

The paper measures the attack success rate of GCG and PAIR jailbreaking attacks (generated on Llama-2-7b-chat) across six target LLMs on AdvBench and MaliciousInstruct. GCG shows extreme model-specificity: it fails on the Llama family (ASR 1.4% on Llama-2-13b-chat, 1.7% on Llama-3.1-8b-instruct) but succeeds on Vicuna-13b-v1.5 (92.3%) and Mistral-7b-instruct (96.1%). PAIR is more consistent, maintaining 50-90% ASR across all targets. The paper attributes this to distributional dependency: the adversarial sequences overfit the source model's parameters and sampling process, so their created high-importance regions do not persist in target models.

Evidence
correlational
Key metric
GCG ASR: 1.4 (Llama-2-13b-chat), 1.7 (Llama-3.1-8b-instruct), 92.3 (Vicuna-13b-v1.5), 96.1 (Mistral-7b-instruct), 27.2 (GPT-4-0613), 43.3 (GPT-O1-Preview) on AdvBench; PAIR ASR: 56.2, 67.3, 79.6, 81.8, 85.6, 72.1 respectively
Caveat
GCG and PAIR were run without their optimal configurations (no ensemble for GCG, no auxiliary judgment LLM for PAIR); optimal settings are reported in Appendix E and show higher but still model-dependent ASR.
Model
Llama 2 / Llama 2 base Llama-2-13B-Chat, Llama 3.1 8B Instruct, Vicuna Vicuna-13b-v1.5, Mistral 7B / Mistral / Mistral 3 7B / Mistral-0.2-7B / Mistral-v0.1 Mistral-7B-Instruct, GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report GPT-4 0613, O1 / OpenAI-o1-preview GPT-O1-Preview
Concepts
Failure mode
Datasets
AdvBench / AdvBench-50 [eval], MaliciousInstruct / i-maliciousinstructions [eval]
Methods
GCG [primary], PAIR [primary]
Related work
GCG [builds-on], PAIR [builds-on]
Related findings
IC-341, IC-342
Extraction
automatic-extraction