Listen to this Post
Introduction: The Moment AI Turned Against Cybersecurity
Artificial intelligence has entered a new and unpredictable phase in cybersecurity. For years, experts warned that attackers would eventually use large language models (LLMs) to automate reconnaissance, vulnerability discovery, and exploitation. That future is no longer theoretical. A newly observed campaign has revealed one of the clearest examples of an AI-powered hacking operation functioning with minimal human intervention.
Security researchers discovered that a Chinese-speaking threat actor created an autonomous hacking framework capable of researching vulnerabilities, selecting targets, testing exploits, and managing attack workflows. However, in an unexpected twist, the same AI automation that helped scale the operation also exposed the attacker’s own infrastructure.
The incident, investigated by Unit 42 researchers, provides a rare look inside a modern AI-assisted cyberattack. The attacker’s autonomous agent accidentally launched a public file server from its working directory, revealing exploit code, configuration files, API credentials, and operational logs. Instead of remaining hidden, the AI system effectively opened a window into the attacker’s entire workflow.
This event represents a major turning point. AI is no longer simply a tool used by security researchers or developers. It has become an operational component in real-world cyber campaigns, capable of making decisions, adapting strategies, and interacting with complex digital environments.
Original Attack Summary: An AI Agent Running the Hacker’s Playbook
The Discovery of an Autonomous Cyber Weapon
Unit 42 researchers identified a threat actor operating under the aliases “knaithе” and “KnYuan.” The attacker built an autonomous hacking environment using DeepSeek integrated with the open-source Hermes Agent framework.
Unlike traditional malware campaigns where operators manually execute commands, this system allowed an AI agent to perform multiple stages of an attack independently.
The AI-driven framework was designed to:
Analyze vulnerabilities.
Search for public exploits.
Identify exposed services.
Select potential targets.
Attempt exploitation.
Record attack progress.
The system was controlled through Telegram, allowing the attacker to remotely interact with the autonomous agent while delegating technical operations to AI.
How the AI Hacker Operated Without Constant Human Control
From Reconnaissance to Exploitation
The most significant aspect of this campaign was the AI agent’s ability to perform a complete attack lifecycle.
During one observed operation, the agent attempted to exploit CVE-2026-33017 affecting Langflow, a vulnerability rated with a critical CVSS score of 9.8.
The AI scanned approximately 84 exposed Langflow installations. After analyzing the environment, it determined that exploitation was unlikely because of configuration restrictions.
Instead of stopping, the AI agent adjusted its strategy and searched for alternative opportunities.
This behavior demonstrates a significant difference between traditional automated scanning tools and modern AI-powered systems. Traditional scanners follow predefined rules, while AI agents can evaluate results, change direction, and prioritize new objectives.
AI-Powered Vulnerability Hunting Through Public Data
The Search for Bigger Attack Opportunities
After abandoning the Langflow exploitation attempt, the AI system began searching GitHub for trending proof-of-concept exploits.
It analyzed publicly available vulnerability research and matched exploit information against exposed online services.
The investigation eventually focused on n8n, an automation workflow platform.
The AI attempted to combine:
CVE-2026-21858 with a CVSS score of 10.0.
CVE-2025-68613 with a CVSS score of 9.9.
These vulnerabilities represented potentially devastating attack opportunities because workflow automation platforms often contain sensitive integrations, credentials, and business processes.
According to exposure intelligence gathered through FOFA, more than 647,000 n8n instances were visible globally. The AI narrowed its search toward Chinese infrastructure and began testing selected targets.
However, authentication requirements prevented successful compromise in the examined cases.
The Biggest Mistake: When AI Exposed the Hacker
Automation Created an Operational Security Failure
The most ironic part of the campaign was not the failed exploits. It was the AI agent’s mistake.
During operation, Hermes Agent executed a command that launched a local HTTP file server from its working directory.
Instead of protecting sensitive files, the AI unintentionally made them accessible.
The exposed data included:
Exploit scripts.
API keys.
Configuration files.
Attack logs.
Internal workflow information.
The attacker attempted to build an advanced autonomous hacking system, but poor AI control mechanisms caused the operation to become visible.
This incident demonstrates a new cybersecurity reality: AI can increase offensive capabilities while simultaneously creating new failure points.
Deep Analysis: Understanding the AI Attack Framework
The Technical Architecture Behind the Campaign
The operation demonstrates how attackers are combining multiple technologies into automated cyber platforms.
The framework consisted of:
Large language models for reasoning.
Agent frameworks for task execution.
Messaging platforms for remote control.
Public vulnerability databases.
Internet scanning services.
Exploit repositories.
The attacker used DeepSeek as the primary AI model but also experimented with several other models, including:
Qwen.
GLM.
Kimi.
MiniMax.
Western AI coding assistants such as Claude Code and Codex were reportedly tested mainly for connectivity checks and exploit development workflows.
Example Commands Observed in Autonomous Workflows
A simplified example of how attackers may automate infrastructure exposure:
python3 -m http.server 8000
This command starts a basic HTTP server.
In legitimate development environments, it can be useful for quickly sharing files. However, when executed inside a hacking workspace containing credentials and exploit materials, it can accidentally expose sensitive information.
Example Attack Automation Workflow
A simplified autonomous attack chain:
1. Discover exposed services
masscan target-range -p80,443
2. Identify vulnerable software
search vulnerability database
3. Download exploit information
git clone exploit-repository
4. Analyze target
AI agent evaluates configuration
5. Attempt exploitation
execute exploit module
6. Record results
save logs and credentials
The danger is not one individual command. The danger is the combination of reasoning, automation, and execution.
An AI agent can now connect multiple steps together at machine speed.
Multiple AI Models Used to Improve Attack Capability
The Rise of Multi-Model Cyber Operations
The threat actor did not rely on a single AI system.
Instead, different models were tested for different purposes.
Some models may provide stronger coding abilities, while others may perform better in reasoning or information analysis.
This suggests a future where attackers could build specialized AI teams:
One AI agent for reconnaissance.
One AI agent for exploit research.
One AI agent for coding payloads.
One AI agent for evasion.
One AI agent for reporting.
Cybersecurity defenders may soon face coordinated AI systems rather than individual attackers.
Attribution Attempts and Evasion Techniques
Trying to Hide Behind AI Infrastructure
The attacker attempted to reduce attribution risks by using proxy infrastructure and disabling certain logging features.
Researchers observed attempts to route activities through:
code.newcli[.]com
The actor also disabled features such as Codex response storage to reduce evidence collection.
However, these precautions failed because the autonomous agent itself created the exposure.
This highlights a new challenge: attackers must now manage not only human mistakes but also AI-generated mistakes.
Confirmed Impact: Hundreds of Targets Scanned
A Large-Scale Automated Operation
Researchers observed the AI framework scanning more than 460 targets.
Although most exploitation attempts failed, at least three compromises were confirmed.
The numbers demonstrate the efficiency advantage provided by autonomous systems.
A human attacker may spend hours manually researching targets. An AI-powered system can perform thousands of evaluations rapidly and continuously.
Even a low success rate becomes dangerous when attack volume increases dramatically.
What This Means for the Future of Cybersecurity
AI Has Officially Entered the Battlefield
This campaign proves that autonomous hacking is moving from research demonstrations into real-world operations.
Attackers are no longer only using AI to write malicious code or generate phishing emails.
They are building systems that can:
Understand environments.
Make tactical decisions.
Search for weaknesses.
Adapt strategies.
Execute attack steps.
The cybersecurity industry must prepare for a world where attackers operate with AI assistants that never sleep and can process massive amounts of information instantly.
Defensive Strategies Against Autonomous AI Attacks
Security Teams Must Change Their Approach
Traditional security methods are not enough against AI-powered threats.
Organizations should focus on:
Continuous Attack Surface Monitoring
Companies need real-time visibility into exposed assets.
Unknown internet-facing systems can become immediate targets for autonomous agents.
Faster Vulnerability Management
Critical vulnerabilities must be patched quickly.
AI attackers can discover and test vulnerabilities faster than traditional security teams can respond.
Strong Configuration Controls
Many successful attacks happen because of weak configurations.
Organizations should:
Remove unnecessary internet exposure.
Protect API keys.
Disable unused services.
Monitor unusual automation behavior.
AI Security Monitoring
Organizations should begin monitoring AI activity itself.
Future security systems will need to detect:
Unauthorized AI agents.
Automated exploitation attempts.
Abnormal API usage.
AI-generated attack patterns.
What Undercode Say:
AI Hackers Are Becoming Real Operational Threats
This incident represents one of the clearest demonstrations of autonomous cyber operations.
The biggest lesson is not that AI can attack systems.
The cybersecurity industry already expected that possibility.
The important discovery is that AI agents can now organize complex attack workflows.
The attacker did not simply ask an AI model for malicious code.
The attacker created an autonomous digital operator.
This changes the entire threat landscape.
Traditional hackers are limited by time, knowledge, and manpower.
AI removes many of those limitations.
A single operator can potentially manage thousands of automated attack attempts.
However, this campaign also revealed a weakness.
Autonomous systems are not perfect.
They can misunderstand instructions.
They can make dangerous assumptions.
They can expose sensitive information.
The same automation that increases offensive power can create operational disasters.
This creates a new cybersecurity battlefield.
Defenders are no longer fighting only human attackers.
They are fighting human-controlled intelligent systems.
AI security will become one of the most important cybersecurity fields.
Organizations will need protection against AI-generated attacks.
They will also need protection against their own AI tools.
Poorly configured AI assistants could become internal security risks.
Attackers will continue improving their autonomous frameworks.
Future AI agents will likely become faster and more specialized.
They may cooperate with each other automatically.
One agent could discover vulnerabilities.
Another could create exploits.
Another could manage persistence.
Another could analyze defensive reactions.
Cybersecurity teams must prepare before these systems become mainstream.
The traditional security model of waiting for attacks is becoming outdated.
Prevention and continuous monitoring will become essential.
AI will become both the weapon and the shield.
The organizations that adapt early will gain a major advantage.
The organizations that ignore AI threats may become easy targets.
This event should be considered an early warning.
Autonomous cyber warfare has already started.
The future of cybersecurity will be defined by AI versus AI.
The next generation of security battles will happen at machine speed.
✅ Confirmed: AI-Powered Attack Framework Discovery
Unit 42 researchers documented an autonomous hacking framework using DeepSeek and Hermes Agent technologies. The exposure of operational files was caused by an AI agent mistake involving a publicly accessible file server.
✅ Confirmed: Vulnerability Research and Target Selection
The campaign involved attempts against vulnerabilities affecting Langflow and n8n environments. Researchers observed automated scanning and exploit-selection behavior.
✅ Confirmed: AI Model Experimentation
The threat actor tested multiple AI models, including DeepSeek, Qwen, GLM, Kimi, and MiniMax, showing the growing popularity of multi-model attack development.
❌ Not Confirmed: Fully Successful Global AI Cyber Warfare
Although this incident proves autonomous attack capabilities exist, it does not mean AI hackers can currently compromise any system automatically. Authentication barriers, security controls, and defensive technologies remain effective.
Prediction
The Future of Autonomous Cyberattacks
(+1) Positive Prediction:
Security companies will accelerate the development of AI-powered defense systems capable of detecting autonomous attackers, analyzing AI-generated behavior, and responding automatically before major damage occurs.
(-1) Negative Prediction:
Attackers will continue improving autonomous hacking frameworks, creating faster and more scalable cyber campaigns that could overwhelm organizations lacking strong security monitoring and rapid patch management.
Final Conclusion: The Hacker’s AI Made the Biggest Mistake
The emergence of this autonomous hacking framework marks a major moment in cybersecurity history. The attacker successfully demonstrated how AI can perform complex cyber operations, but the same technology also exposed the operation itself.
The lesson is clear: AI has entered the cybersecurity battlefield.
The question is no longer whether attackers will use artificial intelligence.
The real question is whether defenders will be prepared when AI-powered attacks become faster, smarter, and more common.
▶️ Related Video (82% Match):
🕵️📝Let’s dive deep and fact‑check.
🎓 Live Courses & Certifications:
Join Undercode Academy for Verified Certifications
🚀 Request a Custom Project:
Secure, high-velocity infrastructure and disruptive technological engineering. Contact our engineering team for high-tier development and proprietary systems:
[email protected]
💎 Smart Architecture | 🛡️ Secure by Design | ⭐ Trusted by Thousands
References:
Reported By: cyberpress.org
Extra Source Hub (Possible Sources for article):
https://www.stackexchange.com
Wikipedia
OpenAi & Undercode AI
Image Source:
Unsplash
Undercode AI DI v2
🔐JOIN OUR CYBER WORLD [ CVE News • HackMonitor • UndercodeNews ]
📢 Follow UndercodeNews & Stay Tuned:
𝕏 formerly Twitter 🐦 | @ Threads | 🔗 Linkedin | 🦋BlueSky | 🐘Mastodon | 📺Youtube
