ResearchBenchmarksCode Backdoor EvaluationFollow-73Insecurity Impact (%)AdL*-76-55.75-35.5-15.25Apr 12, 2026Evaluation ResultsMethodMethodLinksInsecurity Impact (%)Code CE ChangeMMLU Delta Accuracy (%)AdL*Backbone=GemmaBackbone=Gemma2026.04-730—AdL*Backbone=LLaMABackbone=LLaMA2026.04-63.2-0.2—AdLIRA*2026.04-45.10-5.3CBBackbone=GemmaBackbone=Gemma2026.04-11.60—CBBackbone=LLaMABackbone=LLaMA2026.04-8.80.1—CB2026.04-5.30-17.1LIRA*Backbone=GemmaBackbone=Gemma2026.04-4.90—GDBackbone=LLaMABackbone=LLaMA2026.04-3.80.2—GDBackbone=GemmaBackbone=Gemma2026.040.922.6—GD2026.04217.8-2.1