Reproduce “Mechanistic Interpretability of LLM Jailbreaks via Internal Attribution Graphs” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Anupam Wagle, Ifrat Ikhtear Uddin, Chaowei Zhang, Longwei Wang

The authors compare internal computation graphs for benign and attacked prompts to study jailbreaks in Llama-2-7B-chat. They align learned features and measure changes in influence along output-directed paths. Across 30 paired prompts, four attacks succeeded, and path rerouting correlated with success (r = 0.461, p = 0.010). This association comes from a small, single-model study and does not show causation.

Key insights

  • Successful LLM jailbreaks primarily operate by rerouting internal computational pathways rather than simply suppressing safety-aligned features.
  • The magnitude of 'path rerouting' within the internal computation graph is a statistically reliable predictor of jailbreak success, particularly for context-switching attacks.
  • Node-level causal interventions (e.g., zero-ablation) targeting emergent features proved insufficient to mitigate successful jailbreaks, indicating distributed effects and redundancy in LLM internal mechanisms.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.