Test-time policy shaping aligns autonomous cyber defense agents
Original title: Alignment of Autonomous Cyber Defense Agents via Adaptive Policy Shaping
In CybORG, adaptive shaping improved CIA alignment without retraining while keeping defense performance.
The paper frames alignment of reinforcement-learning cyber defense agents as test-time policy shaping, adjusting a pretrained policy with predicted action attributes. An adaptive controller changes the alignment-versus-performance trade-off as priorities shift during an attack. Evaluation uses the CybORG simulator and confidentiality, integrity, and availability objectives. The authors report better alignment without a full retrain, and the code is public.
Unconfirmed: Reported gains are from the simulated CybORG environment, not a production network.
- The approach avoids retraining the decision policy.
- Objectives follow the confidentiality, integrity, and availability triad.
- Code is published in the ITM-Kitware alignment-cage-challenge-2 repository.
- Products
- CybORG
- Organizations
- ITM-Kitware
From the original article
1 min read at arxiv.orgCybersecurity operators increasingly encounter sophisticated and fast-paced attacks, making autonomous cyber defense (ACD) a critical approach for supporting timely and effective decision-making. Reinforcement learning (RL)-based ACD agents have the ability to autonomously observe network states and take actions to defend and recover system assets, making them well suited for ACD.
By Dena Mujtaba · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.