Autonomous AI Weaponization at Scale

Automated Offensive AI Workflows

Security researchers at Unit 42 have uncovered a campaign involving a China-based threat actor who is leveraging the DeepSeek AI model to conduct autonomous offensive operations. By integrating the open-source Hermes Agent, the attacker has transitioned from manual exploitation to a model where AI agents perform reconnaissance and vulnerability assessment with minimal human intervention.

The activity was brought to light after a configuration error in the attacker's infrastructure exposed sensitive internal logs, including API keys and target lists. This discovery provides a rare, transparent view into how modern AI frameworks are being repurposed to accelerate the threat landscape, effectively collapsing the time required for weaponization.

The Mechanics of Autonomous Targeting

The core of the operation utilizes the Hermes Agent, which can be set to a specialized "Yolo" mode. This configuration enables the agent to execute commands and navigate terminal environments without seeking operator permission. The system was designed to interface with FOFA to identify internet-exposed assets, allowing it to parse massive datasets for potential entry points.

While the observed campaign had limited impacts, the workflow confirms a functional, end-to-end autonomous offensive capability.

— Unit 42, Palo Alto Networks

During a session observed in May 2026, the agent autonomously identified vulnerable Langflow servers and searched for exploit code. When the initial targets proved resilient, the AI independently shifted focus, analyzing repositories to select the n8n platform as its next target.

• 84 instances of Langflow were identified as potential targets using FOFA.
• 647,000 instances of the n8n workflow automation platform were scanned.
• More than 460 systems were targeted manually by the actor using various vulnerabilities.
• Three successful compromises were confirmed targeting the Citrix NetScaler vulnerability CVE-2026-3055.

Scaling Speed Through AI Integration

The efficiency gain reported by researchers is significant. Rather than spending dozens of hours on manual target identification, the system performs analysis and scoping in minutes. This speed allows attackers to process large volumes of data and pivot between different vulnerability classes, such as chaining CVE-2026-21858 and CVE-2025-68613, with far less friction than traditional manual methodologies.

While the AI was used for discovery and vulnerability scanning, the actor also maintained a more traditional manual attack strategy for specific infrastructure. This hybrid approach suggests that while autonomous tools are scaling the reconnaissance phase, attackers are still relying on human expertise for more sensitive post-exploitation tasks, such as searching for authentication cookies to hijack sessions.

Implications for Defensive Strategy

The transition toward autonomous agents fundamentally changes the speed at which organizations must identify and patch exposed infrastructure. When attackers utilize autonomous AI to perform tasks that previously consumed hours of manual labor, the window for defenders to detect and remediate vulnerabilities narrows significantly. If an AI can identify, scan, and attempt to exploit a vulnerability in near-real-time, static defense models may prove insufficient. Businesses should evaluate their exposure by assessing whether automated tools, if used by an adversary, would find their public-facing instances easily accessible through search engines like FOFA.