Lazarus attack explained: hijacking LLM output by reprogramming one MLP neuron
DimitrisPapail · x · 2026-11-11
Dimitris Papail breaks down how silly-simple the Lazarus attack is: find a weak neuron, set its in-weights to the "trigger present" direction, let SiLU force 0 when absent, set the out-weights to the unembedding of the token you want screamed into the residual — it goes straight to logits, so that token wins. The accompanying diagram shows the reprogrammed MLP neuron: a trigger-classifier gate, SiLU gating, and a down vector writing the target token direction into the residual stream.
More from Research
- Google launches AlphaGenome 2 to predict how DNA variants affect specific cell types — sundarpichai · 2026-11-12
- New learning rule boosts Hopfield network memory capacity over 10×, per Nature Communications study — bravo_abad · 2026-11-12
- CoRL results show sim-only training now delivers top real-world robot performance — carlosdponx · 2026-11-12
- Google team pushes neural cellular automata to multi-step visual reasoning — bratton · 2026-11-12
- NeoMME: 255x smaller multimodal embeddings while keeping >95% nDCG@10 — antoine_chaffin · 2026-11-12
- Study: Reflective Data Center Roofs Actually Cool Daytime Ground by 1.8-2.2°C — SumitGup · 2026-11-12