Metraipolis › archive › dc_fb6fbd06a764
Circuit Discovery Helps Detect LLM Jailbreaking: A Mechanistic Interpretability Study — arXiv:2608.27504
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.27504
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-17T10:58:10.117Z
- SHA-256 of the archived text
9da755bdf8ab15399620f67d6dc5e7c737bcd6bb5921e7a1e549c7ae605a419a- Stable URL of this page
https://metraipolis.com/documento/dc_fb6fbd06a764
How to cite
Circuit Discovery Helps Detect LLM Jailbreaking: A Mechanistic Interpretability Study — arXiv:2608.27504 — arXiv (en), https://arxiv.org/abs/2608.27504 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_fb6fbd06a764 (Metraipolis archive dc_fb6fbd06a764, sha256 9da755bdf8ab15399620f67d6dc5e7c737bcd6bb5921e7a1e549c7ae605a419a)
BibTeX
@misc{dcfb6fbd06a764,
title = {Circuit Discovery Helps Detect LLM Jailbreaking: A Mechanistic Interpretability Study — arXiv:2608.27504},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.27504},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_fb6fbd06a764; sha256 9da755bdf8ab15399620f67d6dc5e7c737bcd6bb5921e7a1e549c7ae605a419a},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.27504. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Despite extensive safety alignment, large language models (LLMs) remain vulnerable to jailbreak attacks that bypass safeguards to elicit harmful content. While prior work attributes this vulnerability to safety training limitations, the internal mechanisms by which LLMs process adversarial prompts remain poorly understood. We present a mechanistic analysis of the jailbreaking behavior in a large-scale, safety-aligned LLM, focusing on LLaMA-2-7B-chat-hf. Leveraging edge attribution patching and subnetwork probing, we systematically identify computational circuits responsible for generating affirmative responses to jailbreak prompts. Ablating these circuits during the first token prediction can reduce attack success rates by up to 80\%, demonstrating its critical role in safety bypass. Our analysis uncovers key attention heads and MLP pathways that mediate adversarial prompt exploitation, revealing how important tokens propagate through these components to override safety constraints. These findings advance the understanding of adversarial vulnerabilities in aligned LLMs and pave the way for targeted, interpretable defense mechanisms based on mechanistic interpretability. Tema degli esploratori: fiducia digitale e crittografia (retro-catalogo arXiv: annunci del 2026-08-31 in cs.CR). Categorie arXiv: Cryptography and Security (cs.CR). Pagina del lavoro: https://arxiv.org/abs/2608.27504 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-17T10:49:51.624Z dalla pagina https://arxiv.org/catchup/cs.CR/2026-08-31?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).