galwert/anchorrep-data

AnchorRep Attack Artifacts

Evaluation artifacts for the AnchorRep paper (NeurIPS 2026), AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion. These are the adversarial suffixes and training prompts used to evaluate the defense; the trained LoRA defense adapters are released as separate model repositories (see galwert/AnchorRep-*).

Contents

Path	Description
gcg_advbench/advbench_suffixes_all_models.csv	GCG suffixes optimized on 20 source models over AdvBench prompts (2,020 rows), used for the main cross-model transfer evaluation. Columns: model_index, model, example_index, prompt, and the optimized suffix.
gcg_harmbench/gcg_suffixes_<model>_100.json	Out-of-distribution HarmBench GCG suffixes (100 prompts each) for Llama-3, Mistral, Phi-3, Qwen-1.5-14B, and Vicuna.
training_prompts.json	The 30 harmful training prompts (seed 42) used to train the defenders.

Usage

The companion code repository at https://github.com/galwert/AnchorRep consumes these files directly in its evaluation scripts. See the paper appendix for the GCG attack configuration and the full list of source models.

Citation

@inproceedings{anchorrep2026,
  title={AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion},
  author={Wertheizer, Gal and Himelstein, Rom and Peretz, Tomer and Mendelson, Avi},
  booktitle={NeurIPS},
  year={2026}
}