MemJack-Bench — Public MemJack Jailbreak Trajectory Corpus

MemJack-Bench is the MemJack attack
trajectory dataset used in Every Picture Tells a Dangerous Story:
Memory-Augmented Multi-Agent Jailbreak Attacks on VLMs.

Layout

MemJack-Bench/
├── README.md
├── trajectories.jsonl
├── rounds.jsonl
├── splits/{train,val,test,ood}_ids.txt + manifest.json
└── derived/
    ├── dpo_turning_point.jsonl         (1,974 pairs)
    ├── dpo_cross_model.jsonl           (0 pairs: single victim model)
    ├── sft_refusal.jsonl               (~18K)
    ├── sft_multiturn_robust.jsonl      (~920)
    ├── guard_train.jsonl               (55,867)
    ├── defense_pattern_train.jsonl     (55,867)
    └── sft_guard_summary.json

Ethics & licensing

See MemJack-Bench-v2/README.md — identical policy applies. This release is
intended strictly for defensive safety-alignment research and must not be used
to fine-tune or deploy models whose purpose is to produce harmful content.

Citation

@misc{chen2026picturetellsdangerousstory,
      title={Every Picture Tells a Dangerous Story: Memory-Augmented Multi-Agent Jailbreak Attacks on VLMs}, 
      author={Jianhao Chen and Haoyang Chen and Hanjie Zhao and Haozhe Liang and Tieyun Qian},
      year={2026},
      eprint={2604.12616},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2604.12616}, 
}