Inside a Neural Chameleon

Reverse-engineering how a language model hides from activation monitors

A language model that learned to hide# Neural Chameleons showed that models can learn a general skill for altering their neural activity to hide from completely new, unseen activation monitors.[1]Mainly tested using unseen linear probes but evasion also carried to some probe ensembles, MLPs, and attention monitors. How is this even possible? This seems like a bizarre meta-skill to be capable of. McGuinness et al. describes that: “When triggered, the model pushes its activations into a narrow, low-dimensional subspace, effectively steering them away from the regions a probe would flag as…

saved by

• Cheikh Fiteni
• Jackson Mowatt Gok
• Eva Seigneurbieux

related reading

• [2512.11949] Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitorsarxiv.org
• Circuit Tracing: Revealing Computational Graphs in Language Modelstransformer-circuits.pub
• Neural Chameleons: LLMs Can Learn to Evade Activation Monitorsneuralchameleons.com
• A Mathematical Framework for Transformer Circuitstransformer-circuits.pub
• Verbalizable Representations Form a Global Workspace in Language Modelstransformer-circuits.pub
• In-context Learning and Induction Headstransformer-circuits.pub
• Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activationstransformer-circuits.pub
• Tracing the thoughts of a large language model \ Anthropicanthropic.com
• Natural Language Autoencoders \ Anthropicanthropic.com
• When Chain of Thought is Necessary, Language Models Struggle to Evade Monitorsarxiv.org
• [2608.21664] Measuring Activation Control in Large Language Modelsarxiv.org
• Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Dataalignment.anthropic.com