CircuitBreak: Pruning a Sparse Safety Circuit of Neurons

Summary

CircuitBreak is a neuron-pruning attack that treats LLM refusal as a small circuit of feed-forward neurons whose joint ablation matters, not each neuron alone. It greedily removes a sparse set of 29 to 150 neurons (0.0066% to 0.0388% of the MLP) without twin prompts. On ten instruction-tuned models, attack success rises to about 88% on several Llama and Qwen models, as judged by Llama Guard 3 8B on StrongREJECT.

Related items

• MediumAI agents aren’t ‘going rogue.’ Frontier labs are skipping security basics.Same vendor · CSO Online
• LowAnthropic warns self-replicating AI could spread to your smart devicesSame vendor · BleepingComputer
• InfoShopify opens its catalog up to more AI agentsSame vendor · The Verge (AI)
• InfoCompanies rushed in on AI. Now some are pulling backSame vendor · The Guardian Technology
• InfoTop GenAI security resources — October 2026Same vendor · Adversa AI Blog