Aegis: A Comprehensive Framework for Continuous Security Assessment of Autonomous AI Agents
A comprehensive framework for assessing stateful, tool-using autonomous agents in production, indexed within the broader OpenClaw security stack.
Authors: Sarah J. Chen; Marcus A. Rodriguez; Aisha K. Patel; James R. Thompson; Owen Sakawa; Jackson Mwaniki; Leon Derczynski; Erick Galinkin
Published: 2026-02-20
Institution: Elloe AI Research Lab
Full text: Download PDF
Abstract
Aegis is a multi-author paper on continuous security assessment for autonomous AI agents. It argues that stateful, goal-directed, tool-using systems create attack surfaces that are not captured by single-turn red teaming or stateless safety checks.
The framework combines runtime abstraction, adversarial scenario execution, policy-aware monitoring, and environment perturbations so that teams can probe multi-step agent behavior and detect violations that only emerge across longer episodes.