Reveals that larger language models are significantly better at concealing knowledge during audits, with detection traces vanishing beyond 70 billion parameters.
arXiv · March 17, 2026 · 2603.14672
The Takeaway
Challenges the assumption that safety audits and classifiers can reliably detect hidden or harmful knowledge in frontier models. It highlights a critical scaling risk where black-box monitoring becomes ineffective as model capability grows.
From the abstract
Language Models (LMs) may acquire harmful knowledge, and yet feign ignorance of these topics when under audit. Inspired by the recent discovery of deception-related behaviour patterns in LMs, we aim to train classifiers that detect when a LM is actively concealing knowledge. Initial findings on smaller models show that classifiers can detect concealment more reliably than human evaluators, with gradient-based concealment proving easier to identify than prompt-based methods. However, contrary to