Risk-Adaptive Continuous Attestation for AI Agents

Abstract

Autonomous AI agents increasingly execute real-world actions through tool calls, including file operations, credential access, and financial transactions. We propose Risk-Adaptive Continuous Attestation (RACA), a framework that dynamically determines when an AI agent must undergo fresh attestation before executing an action, based on the assessed risk level of that action. Low-risk actions reuse existing trust evidence while high and critical risk actions trigger mandatory fresh attestation. Evaluation across 1,200 trials demonstrates RACA blocks 100% of high and critical risk attacks while reducing attestation overhead by 49.5% and maintaining sub-millisecond decision latency of 0.00099 ms per call.