Avaliação da Herança de Propriedades Adversariais em Grandes Modelos de Linguagem com Restrição de Contexto e Destilação de Conhecimento Latente

• Gabriel de A. Campos UFF
• Martin Andreoni UFF
• Diogo M. F. Mattos UFF

DOI: https://doi.org/10.5753/semish.2026.23896

Resumo

Este trabalho propõe o arcabouço Prompt Inheritance, voltado à geração de ataques de jailbreaking em Grandes Modelos de Linguagem (LLMs) via destilação de conhecimento latente. A metodologia avalia a herança de propriedades adversariais comparando a manipulação do espaço de embeddings com a exposição a contextos restritos. Experimentos controlados sem ajuste de parâmetros indicam taxas de sucesso de até 61% com prompts concisos. Os resultados demonstram que a proposta supera abordagens da literatura em eficiência, exigindo significativamente menos recursos computacionais e viabilizando ataques escaláveis em hardware acessível.

Referências

Chu, J., Liu, Y., Yang, Z., Shen, X., Backes, M., and Zhang, Y. (2025). Jailbreakradar: Comprehensive assessment of jailbreak attacks against llms.

Delchambre, L. (2014). Weighted principal component analysis: a weighted covariance eigende-composition approach. Monthly Notices of the Royal Astronomical Society.

Huang, D., Shah, A., Araujo, A., Wagner, D., and Sitawarin, C. (2025). Stronger universal and transferable attacks by suppressing refusals. In Chiruzzo, L., Ritter, A., and Wang, L., editors, Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 5850–5876, Albuquerque, New Mexico. Association for Computational Linguistics.

Li, X., Ning, Y., Bao, Z., Xu, M., Chen, J., and Qian, T. (2025). CAVGAN: Unifying jailbreak and defense of LLMs via generative adversarial attacks on their internal representations. In Che, W., Nabende, J., Shutova, E., and Pilehvar, M. T., editors, Findings of the Association for Computational Linguistics: ACL 2025, pages 6664–6678, Vienna, Austria. Association for Computational Linguistics.

Liu, X., Li, P., Suh, E., Vorobeychik, Y., Mao, Z., Jha, S., McDaniel, P., Sun, H., Li, B., and Xiao, C. (2024). AutoDAN-Turbo: A lightweight and scalable framework for automated jailbreaking. arXiv preprint arXiv:2402.16439.

Lyu, Z. and Yuan, M. (2025). Large-dimensional factor analysis with weighted pca. arXiv preprint.

Mazeika, M. et al. (2024). Harmbench: A standardized evaluation framework for automated red teaming and robust refusal.

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E. H., Le, Q. V., and Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems, volume 35, pages 24824–24837. Curran Associates, Inc.

Zhao, Z., Ma, Y., Jha, S., Pavone, M., McDaniel, P., and Xiao, C. (2025). Armor: Aligning secure and safe large language models via meticulous reasoning.

Zhu, S., Zhang, R., et al. (2024). AutoDAN: Interpretable gradient-based adversarial attacks on large language models. arXiv preprint.

Zou, A., Phan, L., Chen, S., Campbell, J., Guo, P., Rens, R., Pan, A., Yin, X., Mazeika, M., Dombrowski, A.-K., et al. (2023a). Representation engineering: A top-down approach to AI transparency. arXiv preprint.

Zou, A., Wang, Z., Carlini, N., Nasr, M., Kolter, J. Z., and Fredrikson, M. (2023b). Universal and transferable adversarial attacks on aligned language models. arXiv preprint.