Google Report: Hackers Use AI Agents to Automate Attacks and Exploit Security Patches Within Hours

iconTechFlow
Share
AI summary iconSummary
Liquidity and crypto markets face new risks as Google reports hackers are using AI agents to automate attacks. A recent breach exploited a newly patched PHP vulnerability using an open-source framework, targeting French government and commercial websites. The report outlines five AI-driven trends, including LLMJacking and attacks on crypto firms. CFT agencies worldwide now face challenges in tracking rapidly evolving threats.

Author: Google Threat Intelligence Group (GTIG)

Compiled by Deep潮 TechFlow

Shenchao Overview: The balance between cyber offense and defense is being fundamentally disrupted by the automation enabled by generative AI. According to the latest global tracking report released by Google Threat Intelligence Group (GTIG), hackers have officially moved beyond the initial stage of simply querying AI and have fully upgraded to end-to-end automated cyber weapons driven by coordinated multi-agent systems.

In a recently disclosed real-world intrusion, attackers deployed an automated vulnerability discovery pipeline based on an open-source multi-agent framework, directly transforming the official newly released PHP security patch into weaponized exploit code and launching automated penetration and data theft against French government and commercial websites within a very short time.

This article provides a comprehensive review of the evolution timeline of multi-agent infiltration frameworks, the black market ecosystem for reselling cloud-based large model computing power (LLMJacking), and how frontline defenders are rebuilding automated security defenses in the age of agents.

Executive Summary: Attackers Have Officially Entered the Era of Autonomous Multi-Agent Systems

As hacker organizations accelerate from basic prompt-based queries to deploying fully autonomous agent workflows, the tactical use of generative AI in real-world cyberattacks has reached a major turning point.

Based on Mandiant’s frontline incident response experience and global telemetry data, GTIG has identified the five key evolutionary trends in global AI threats over the past quarter:

1. Shift to autonomous workflows and multi-agent frameworks: Attackers are no longer using large language models (LLMs) merely as productivity aids; they are now integrating fully autonomous execution agents that coordinate concurrent tasks such as reconnaissance, code auditing, and vulnerability exploitation across multiple models.

2. Automated Vulnerability Discovery and Patch Reverse Engineering: Hackers are leveraging large models to systematically analyze newly released security patches, reverse-engineering the vulnerability trigger logic within hours, completely eliminating the defense team’s patch window.

3. Compute Hijacking and AI Credential Monetization (LLMJacking): Driven by profit, cybercriminals extensively infiltrate enterprise cloud environments, stealing high-value LLM API keys and compute quotas, then selling them on the dark web at discounted rates through “model shuffling mixers” to cash out.

4. End-to-end automation of false perception operations (IO): State-level hackers and commercial influence warfare contractors have deeply integrated large models into troll networks, enabling fully automated fake account registration, realistic profile photo generation, and personalized, targeted information dissemination.

5. Targeted infiltration targeting cryptocurrencies and high-tech companies: North Korea and hackers with specific backgrounds widely abuse AI-generated resumes and voice spoofing tools to impersonate job candidates, generate real-time answers during technical interviews, and carry out deep supply chain poisoning after employment.

Multi-agent framework in network intrusions

In previous tracking studies, GTIG documented mostly isolated, small-scale experiments by hackers testing LLMs; however, the current engineered adoption of autonomous multi-agent frameworks by attackers marks the formal escalation of cyber warfare into scaled, autonomous collaboration.

The core paradigm of this attack framework involves integrating the classic OODA (Observe, Orient, Decide, Act) loop, with the master agent assigning tasks and specialized sub-agents concurrently performing network topology scanning, fingerprinting, script generation, and lateral movement.

Diagram: Collaborative Execution Process of a Multi-Agent LLM Attack Framework

In addition to the recent penetration testing tool ARTEX, which has garnered significant attention in the security community, GTIG has independently captured multiple open-source offensive agents directly used by attackers for frontline intrusions:

• CAIRN: An open-source multi-agent penetration testing framework based on the OODA loop, capable of autonomously scanning unprotected internal network hosts and dynamically generating exploits without any human intervention.

• CyberStrikeAI: A native offensive framework designed for LLMs, featuring built-in sub-agents specialized in privilege persistence and credential theft.

• HexStrike AI: An active offensive tool deeply integrated with the Model Context Protocol (MCP), seamlessly bridging various hacker penetration command-line tools to enterprise-grade large models.

• PentAGI: An automated tool designed to significantly lower the technical barrier to cyberattacks, enabling even novice script kiddies with no understanding of underlying vulnerabilities to issue complex intrusion commands.

Figure: Roadmap for the Evolution of LLM Maturity in Real-World Cyber Attack and Defense Scenarios

Using agents for vulnerability research and development: turning patches into attacks

The most alarming real-world threat posed by multi-agent technology is its ability to reduce the time required for vulnerability discovery and weaponization from days to minutes.

Attackers automatically scrape official open-source patch code from GitHub and feed it to a code analysis agent, which uses the model to compare code diffs in real time and precisely identify underlying logic vulnerabilities that have not yet been widely patched.

Chart: Comparison of the number of N-Day and 0-Day vulnerabilities exploited by hackers in the wild since 2024

In the latest identified real-world security incident, GTIG tracked down and captured an attack staging server publicly exposed on the internet. The server hosted an automated vulnerability discovery framework based on LLMs, with Russian-language development logs prominently included in the code comments:

LLM API key for automated exploit exploitation (exploit_agent pilot project)
You must manually enter the key into the environment variables; never input it into public chat channels. Both configuration lines are required.

The hacker group (codenamed Emad) leveraged this framework to perform fully automated reverse compilation targeting recently disclosed PHP core vulnerabilities (CVE-2026-109618 and CVE-2026-109619), generating live attack payloads within hours of the patch release. Subsequently, the automated cluster launched large-scale, indiscriminate credential-stuffing attacks against multiple European government and commercial web portals, including those in France, successfully exfiltrating sensitive underlying data.

Compute Hijacking (LLMJacking): How Hackers Monetize Stolen Cloud-Based Large Models

As enterprises fully migrate to the cloud and generative AI infrastructure becomes widespread, the risks of stolen cloud credentials are rapidly extending to model computing power.

Over the past several months, Mandiant’s frontline team has handled multiple new cybercrime cases known as “LLMJacking” (Large Model Compute Hijacking). After infiltrating enterprises’ AWS or Google Cloud environments, attackers no longer prioritize deploying mining malware; instead, they quickly activate the victim’s high-end large model API access permissions.

Chart: Victim company's cloud experienced a catastrophic spike in large model API calls due to LLMJacking

Hackers used stolen administrator credentials to create hidden IAM roles and service accounts within the enterprise, then connected them to the dark web's "LLM Mixer gateway."

On the dark web, numerous illegal resale platforms similar to APIMart have emerged. Hackers are publicly selling stolen legitimate commercial large model APIs to other cybercriminals at prices仅为 20% to 30% of the official rate. Affected companies not only face the risk of core data leakage but also incur monthly bills amounting to hundreds of thousands of dollars due to the hackers’ excessive API usage.

Figure: Real-time pricing on the dark web platform APIMart for stolen commercial large model APIs

Large models are fully integrated into disinformation and information warfare.

In the realm of cognitive warfare and online bot armies, LLMs have become the key lever disrupting the traditional cost structures of bot farms.

A commercialized "Influence-as-a-Service" criminal gang has developed a turnkey automated system for manipulating public opinion. The system automatically matches fake accounts with the speech patterns and local slang of targeted audience segments, and even deliberately introduces slight lens noise, blur, and angular tilt when generating images via diffusion models—mimicking casual smartphone snapshots so convincingly that it easily bypasses AI detection algorithms on major social media platforms.

Meanwhile, to circumvent IP bans on social platforms, the manipulation group routes traffic through a global network of residential proxies, combined with bulk automated one-time phone number verification services, to build a fully decentralized, instantly renewable bot army network.

Google's Active Security Defense and Ecosystem Hardening

Facing the growing challenge of agent weaponization, Google is fully leveraging frontline threat intelligence to reinforce the underlying security alignment of its models, building a multi-layered proactive defense architecture:

1. Real-time Model Enhancement and Immunity to Adversarial Examples: The latest hacker prompts and multi-agent attack logic captured in the wild by Mandiant are immediately incorporated as adversarial training data for Google DeepMind and the security team, blocking the model’s response to malicious code and penetration commands at the foundational level.

2. Google AI Threat Tracker Threat Monitoring Platform: Provides real-time monitoring of malicious agent behavior and alerts for abnormal usage volumes for enterprise customers worldwide, blocking unauthorized LLMJacking traffic within seconds.

3. Security Operations Automation (SecOps): Deploy specialized agent behavior monitoring models in the Cloud Security Center to use AI against AI, precisely identifying latent signatures of autonomous malware through analysis of network communication topologies and system call sequences.

Disclaimer: The information on this page may have been obtained from third parties and does not necessarily reflect the views or opinions of KuCoin. This content is provided for general informational purposes only, without any representation or warranty of any kind, nor shall it be construed as financial or investment advice. KuCoin shall not be liable for any errors or omissions, or for any outcomes resulting from the use of this information. Investments in digital assets can be risky. Please carefully evaluate the risks of a product and your risk tolerance based on your own financial circumstances. For more information, please refer to our Terms of Use and Risk Disclosure.