Agentic large language models (LLMs) are models that can reason, plan, and execute tools over unstructured data. These abilities enable transformative applications in domains spanning personal assistants, financial, and legal work. However, such capabilities also raise privacy concerns, as agents may handle sensitive information while interacting with external tools.

The paper 'NeuroFilter: Activation-Based Guardrails for Privacy-Conscious LLM Agents' introduces a guardrail method based on internal activations. The approach, called NeuroFilter, is designed to address privacy in LLM agents by filtering or constraining their behavior. The paper is posted on arXiv under identifier 2601.14660, with the latest version being a replacement (v3).