What It Means and How to Counter Autonomous Offensive Cyber Agents

For years, we’ve anticipated this day. With the release of Anthropic’s landmark report (detailing the disruption of a cyberespionage operation orchestrated by AI agents with minimal human intervention), the reality of autonomous offensive cyber agents has moved from speculation to an active, machine-speed threat. The report covers their internal identification and analysis of artifacts from the GTG-1002 campaign, which was conducted against over 30 different enterprise targets. This event is independently being tracked in the AI Incident Database as incident 1263. To have a successful defense in the age of AI, we need an immediate shift from human-led, reactive security to a proactive, machine-driven security paradigm.

The **GTG-1002 campaign **is the first open report of an AI agent, powered by Claude Code, targeting multiple enterprise environments. Using Claude Code as the primary orchestration framework, the agent was effective in all key phases of the attack:

• Mapping attack surfaces without human guidance.
• Exploit vulnerabilities using custom code generation.
• Moving laterally by autonomously harvesting and testing credentials.
• Conducting an intelligence analysis to identify and prioritize high-value data, rather than just exfiltrating raw dumps.

It was a watershed moment for several key reasons:

• Stealth – Traffic analysis of the inputs and outputs to Claude Code were the initial indicators of this attack, however, the attack was only observable in aggregate.
• Self-Configuration – The agent autonomously adapted its attack strategy to achieve actions on an objective.
• **Machine-Speed – **The agent both orchestrated AND executed the campaign across all attack vectors.
• Autonomous Context and Persistence – Using structured markdown files, the execution agent maintained a persistent state of the attack, providing context and autonomous continuity between distributed sub-actions and attack phases.

This campaign, executed at “multiple operations per second,” marks the end of the necessity for the "human-in-the-loop” attacker and the arrival of the "human-on-the-loop" supervisor. Transitions between attack phases were controlled by the human to validate sufficient completion of the current phase before progressing. It was a thin layer of supervisory human control. With the whiplash pace of AI, defenders should anticipate the necessity of any human control to fade.

In the reported attack campaign, “commodity tools” were leveraged by the threat actor, which at first glance, may not seem particularly novel. However, the autonomous orchestration of these tools across multiple attack phases by Claude Code, using Model Context Protocol (MCP) servers, represents a sophisticated technical advancement in offensive agents. Critically, this method improved more than just the speed of the attack, it also introduced the concept of autonomy with negligible human supervision, supporting dynamic and contextual reasoning in attack path planning across multiple target systems (even beyond typical human analyses, particularly for non-intuitive/interpretable event logging). Custom tools can bring very targeted actions within the same or similar offensive agent architectures, and defenders should be ready for this inevitable evolution.

We Need Agents to Fight Agents

With the debut of real-world offensive agent operations, it is now crystal clear: Defenders cannot combat autonomous, offensive AI with manual, static human driven security operations. Defenses must blend machine-speed responses with on-the-fly adaptability to maintain effectiveness against the self-optimizing campaigns now being observed. The pivot to autonomous agent-driven security operations will require transforming many elements of the traditional security operations lifecycle. All stages from preparation to response processes need to be resilient and robust to changes in adversary speed, stealth, evasion, orchestration frameworks and indicators of compromise.

Meeting the Challenges of Machine-Speed Defense Head-On

A new defense paradigm must be adopted to effectively combat AI attacks that are both orchestrated AND executed beyond human reaction time. To transform security operations and outpace AI-driven threats, organizations need to employ the following core principles:

• Precision of AI for Cybersecurity: Operating at machine speed requires precision and accuracy. Security systems must be capable of ingesting the right data, at the right time, and understanding the system context to detect and block threats in real-time, thwarting AI-generated attacks without generating erroneous alerts. Producing false positives is problematic at human speeds, and the problem compounds at machine speed.
• Proactive Cybersecurity for AI Systems: We must safeguard AI systems with real-time security solutions, preventing the models and applications from being directly or indirectly co-opted for malicious use. This demands a deep and continuous understanding of how AI agents might be abused via their application interfaces, permissions, provenance, identity and wider interactions across organizations.
• Transform Visibility into Observability: Visibility only encompasses a direct presence or absence. Observability is the combination of visibility plus some degree of cognitive and contextual reasoning. The visibility of a traffic sign does not guarantee a driver will observe and respond to it. The GTG-1002 attack evaded detection by splitting and distributing small, seemingly benign fragments of the full campaign across numerous sessions. The requests were visible, but the scope of the malicious campaign was not observed from the isolated requests. To identify and help stop such techniques, defenses need distributed observability, which can only be achieved from context-aware agents that understand the nature and impact of disparate events and can disrupt such attacks when they are identified.
• Agentic Security Operations: As an industry, we must also acknowledge the difference between autonomous and automated systems. The industry has been integrating elements of automation for years. Scripting, decision trees and playbooks are mechanisms for speeding up the response in specific context, but do not necessarily generalize or work across different phases. If the attacker is using an agentic system for 90% of the attack lifecycle, security operations centers (SOCs) must also implement an agentic system for 90% of their triage, investigation, remediation and threat hunting workflows. This must be the rule, rather than the exception. By combining observability with dynamic AI agents capable of coordinated decision making and task execution, SOCs can deliver proactive autonomous protection at scale.

The Future Is Now. Are You Ready?

The GTG-1002 campaign is a clear signal that offensive AI agents are being used in the wild. The adoption of AI agents by threat actors will accelerate and demand a decisive transformation of defensive security operations to include agent orchestration tools customized to respond to the uniqueness of offensive AI agents.

At Palo Alto Networks, our platformization strategy was built precisely for this moment. This interconnectivity between tools and systems transforms visibility into observability necessary for AI agent orchestration.

In light of GTG-1002, there is an unequivocal need for the security community to accelerate the pivot from automated to autonomous security operations. AI agents can quickly find and exploit vulnerabilities, moving stealthily across the attack chain. We must shift from human-led, reactive defense to fast, proactive machine-driven security to ensure cyber resilience in the age of AI.

Are you ready? Learn about securing AI agents and how to create a trustworthy AI ecosystem.

Key Takeaways

• Autonomous Orchestration and Execution: The GTG-1002 campaign was a watershed event because the AI agent, powered by Claude Code, autonomously orchestrated and executed all key phases of the attack, from mapping surfaces and exploiting vulnerabilities to moving laterally and conducting intelligence analysis at machine speed.
• Shift to Machine-Driven Security Paradigm: The emergence of autonomous offensive cyber agents, as demonstrated by the GTG-1002 campaign, demands an immediate pivot from human-led, reactive security to a proactive, machine-driven security defense model.
• Distributed Observability is Essential to Agentic Defenses: To counter new attack techniques like GTG-1002, which evade detection by splitting the campaign into small, distributed, and seemingly benign fragments, defenses must adopt distributed observability to connect disparate events using context-aware agents.

Further Reading:

• Lohn, Andrew, et al. "Autonomous Cyber Defense." A roadmap from lab to ops. Online. Centre for Emerging Technology and Security (CETaS) at The Alan Turing Institute (2023).
• Principles for the Secure Integration of Artificial Intelligence in Operational Technology