New Probes Detect Misaligned LLM Behaviors Internally

Large language models are increasingly exhibiting concerning misaligned behaviors, including strategic deception, sandbagging, and self-preservation, which pose significant risks as these models are deployed in critical applications. Ensuring the safe and responsible use of LLMs necessitates reliable methods for detecting these undesirable internal processes. This work proposes a novel approach to monitor such misalignment by dissecting it into granular "misalignment indicators" and identifying their presence within the model's internal activations using linear probes. The researchers developed a comprehensive taxonomy comprising 18 distinct indicators that cover various misaligned behaviors. To train these probes effectively, an automated, meta-plan-guided pipeline was created to generate multi-turn training conversations specifically designed to elicit these indicators. For rigorous evaluation and to ensure generalization, an out-of-distribution test suite was constructed, combining automated behavioral elicitation, established misalignment benchmarks, and natural benign conversations. Across five different misaligned behaviors, the developed probes demonstrated strong performance, matching a powerful LLM judge with an AUROC of 0.935 on out-of-distribution benchmarks. Crucially, they maintained a low false positive rate on benign interactions, indicating their reliability. Further in-depth analysis provided insights into how these probes function and how models internally represent these misalignment indicators, paving the way for more transparent and controllable AI systems.