Robotics Security and Privacy Initiative logo

When LLMs Autonomously Attack - July 24, 2025

CAI's Co-Directors show how LLMs can be taught to autonomously plan and execute real-world cyberattacks against enterprise-grade network environments—and why this matters for future defenses.

Research Thrusts

Creating the scientific foundations of autonomous cyber systems will require a new, interdisciplinary, systems-thinking approach. The initiative will advance the theory and practice of cyber autonomy along four interconnected research thrusts:

Foundations for Autonomous Attack and Defense Capabilities

Our goal in this thrust is to develop foundational AI and agentic capabilities for autonomous offense and defense. Today, developing novel system-level attack and defense systems is challenging because they are built with low-level tools. We need higher-level system design, algorithmic strategy, and analytics abstractions to accelerate the design and implementation of future autonomous attack and defense capabilities. We also envision novel security-centric AI capabilities based on foundation models, in-context learning, reinforcement learning, and game theory.

Human-AI Collaboration for Cyber Autonomy

As defenses become increasingly autonomous, we need to develop a better understanding of the real-world constraints of operators so that we can design defense systems they will trust and deploy. The Cyber Autonomy Initiative will explore several aspects of Human-AI collaboration in an autonomous world, including debugging attack & defense systems and Human-AI collaborative pentesting.

Infographic image featuring CAI's research vision

Systems Support for Autonomous Operations

We need systems advances and experimental work to inform the autonomous operations envisioned above. The Cyber Autonomy Initiative will explore several aspects of systems advances, including emulation systems and datasets for training models; sandboxing, staging, and verification systems; scalable telemetry and log analytics; and novel software-defined data-plane and control-plane capabilities.

Cyber Autonomy Arena

We will create system-level, competitive evaluations of attackers against defenders on realistic networked systems to better understand the competitive interplay between attack and defense systems in realistic contexts. We will address key challenges in scenario realism and diversity, accelerating the design and contribution of novel attack and defense systems, and privacy-preserving mechanisms for incentivizing community participation.

Meet our co-directors

Lujo Bauer

Lujo Bauer

Professor
Electrical and Computer Engineering, Software and Societal Systems

View bio

Vyas Sekar

Vyas Sekar

Tan Family Professor
Electrical and Computer Engineering

View bio

Become a sponsor

Interested in getting involved or sponsoring this initiative? Contact Michael Lisanti, Senior Director of Partnerships, at 412-268-1870 or mlisanti@andrew.cmu.edu.