top of page

Google Exposes the Rise of Agentic Cybercrime: ARTEX, CAIRN, PentAGI, and the New Threat Landscape

1 day ago
13 min read
Further Reading / External References
GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold

https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold

Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches

https://thehackernews.com/2026/11/google-finds-hackers-using-ai-agents-to.html

Artificial intelligence is transforming cybersecurity, but the technology is also changing how cybercriminals and state-sponsored threat actors conduct attacks. What once required experienced operators to research vulnerabilities, write malicious code, manage infrastructure, and coordinate separate stages of an intrusion can increasingly be automated through large language models (LLMs) and autonomous AI agents.

New findings from Google Threat Intelligence Group (GTIG) and Mandiant reveal that adversarial AI activity has moved beyond experimental prompting into operational deployment. Threat actors are using multi-agent frameworks, automated vulnerability research pipelines, AI-assisted phishing platforms, and compromised cloud infrastructure to increase the speed and scale of cyber operations.


The shift is particularly significant because AI agents can connect multiple technical capabilities into continuous workflows. Instead of simply generating instructions or suggesting code, an agentic system can inspect an environment, select tools, execute commands, evaluate results, and determine subsequent actions.

The emerging threat is not that AI has independently replaced human hackers. Rather, it is that human operators can supervise automated systems capable of performing more tasks, processing more targets, and compressing attack timelines.

This development creates a new security challenge for governments, enterprises, cloud providers, and technology companies. Traditional defenses must increasingly account for attacks that operate at machine speed, exploit legitimate credentials, manipulate software development workflows, and abuse the same AI infrastructure organizations depend on for productivity.


The Rise of Autonomous AI Agents in Cybercrime

The latest threat intelligence identifies a growing shift toward open-source, multi-agent offensive frameworks. These systems coordinate specialized agents to perform different stages of security testing and, when misused, unauthorized intrusion activity.

Among the frameworks identified in real-world threat activity are ARTEX, CAIRN, CyberStrikeAI, and PentAGI. GTIG has also observed experimentation involving HexStrike AI, Agent Zero, and other custom-built orchestration systems.

Their importance lies in coordination. A traditional automated security tool usually performs a defined function, such as scanning a server or checking a vulnerability. A multi-agent framework can coordinate several functions through a shared operational process.

For example, one component may collect information about a target environment, another may analyze potential weaknesses, and another may evaluate available tools and execution results. The system can then adjust its approach based on feedback.

Several technical characteristics make these frameworks particularly relevant to defenders:

  • Autonomous task coordination: Multiple agents divide complex operations into smaller activities.

  • Dynamic tool selection: LLMs can determine which available utilities are relevant to a particular task.

  • Adaptive execution: Results from previous actions influence subsequent decisions.

  • On-demand code generation: Models can produce scripts and temporary payloads for specific operational requirements.

  • Continuous feedback: Agents can evaluate execution outcomes and modify their approach.

CAIRN uses an Observe, Orient, Decide, Act architecture, commonly known as the OODA loop. Its specialized agents collaborate through a shared information structure to progress toward an operational objective. GTIG observed its use in intrusion activity involving targets in sectors including aerospace in the United States.

CyberStrikeAI takes a different approach, combining a Go-based backend that provides command-and-control and orchestration functionality with commercial and open-weight models, more than 100 security tools, YAML attack recipes, and custom scripts.


HexStrike AI integrates with the Model Context Protocol (MCP), allowing LLM clients to interact with more than 150 offensive security utilities through specialized functions and autonomous sub-agents.

PentAGI is designed to reduce the technical complexity involved in coordinating offensive operations. GTIG tracked a customized platform based on its open-source architecture conducting reconnaissance and vulnerability scanning against organizations across Asia between late August and early October 2026.

These developments indicate that the accessibility of advanced offensive capabilities is increasing. However, automation does not eliminate the need for operational judgment, and the reliability of an attack still depends on the underlying vulnerabilities, permissions, infrastructure, and execution environment.


AI-Powered Vulnerability Research Is Shrinking the Exploitation Window

One of the most consequential developments is the use of LLMs to automate vulnerability research and exploit development.

Historically, discovering and weaponizing a software vulnerability required significant technical expertise. Researchers needed to understand source code, identify weaknesses, examine patches, construct proof-of-concept exploits, and validate their findings.

AI-assisted systems can accelerate several of these activities.

A particularly important technique is differential patch analysis. When software developers release a security update, the changes between vulnerable and patched versions can reveal which part of the application contained a weakness.

LLMs can examine these changes, interpret the surrounding code, identify potentially unsafe behavior, and assist with vulnerability analysis. When connected to automated testing environments, agentic systems can also help evaluate whether a suspected weakness is exploitable.

This creates a serious challenge for organizations that delay security updates.

A vulnerability may become easier to understand once its patch is publicly available. Attackers can use the information exposed by the correction to investigate systems that have not yet been updated.


The distinction between zero-day and n-day vulnerabilities is therefore increasingly important.

Vulnerability category

Meaning

Security implication

Zero-day

A vulnerability exploited before an effective patch is publicly available

Defenders may have limited time and information to respond

N-day

A publicly known vulnerability for which a patch or mitigation exists

Unpatched systems remain exposed after disclosure

Patch-derived vulnerability

A weakness identified through analysis of software changes introduced by a security update

Attackers can use patch information to accelerate exploit development

GTIG reported that overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploitation, meanwhile, increased from an average of eight incidents per month in 2025 to 11 per month in 2026, before reaching 22 in August.

The figures suggest that AI-assisted exploitation is not limited to discovering previously unknown weaknesses. It can also increase the speed at which attackers identify and exploit vulnerabilities that organizations already have an opportunity to fix.

The exploit_agent Framework Demonstrates the Next Stage

GTIG identified an exposed staging server hosting a multi-agent vulnerability research system called exploit_agent.

The framework integrated multiple frontier and locally hosted LLMs to automate continuous analysis of popular web application software. It compared source code before and after security patches, examined potential weaknesses, developed exploit code, and executed tests without requiring a human operator at every stage.

Investigators also identified scripts associated with post-compromise activities, including internal reconnaissance, brute-force attempts, webshell installation, and removal of competing attackers' webshells.

Telemetry suggested that the operator had a particular interest in French websites associated with media, entertainment, and education.

This case illustrates how AI can connect vulnerability research with operational intrusion activities. The important security implication is the reduction of friction between discovering a weakness and attempting to exploit it.


Real-World Exploitation Shows How Quickly AI Can Accelerate Attacks

The risks are not limited to experimental frameworks.

GTIG highlighted exploitation activity involving two PaperCut vulnerabilities identified in late August 2026:

  • CVE-2026-81578, an authentication bypass vulnerability.

  • CVE-2026-82078, an unsafe dynamic class loading vulnerability.

When chained together, the flaws could enable unauthenticated remote code execution.

Security researchers and Huntress reportedly generated working exploits within a day of disclosure using LLM assistance. GreyNoise subsequently identified a campaign in which an attacker appeared to use AI to develop and test exploits, assemble targets, and launch widespread exploitation activity against hundreds of organizations.

The attacker reportedly expanded an initial foothold to domain administrator privileges within minutes.


Another example involved PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619. LLM-enabled analysis helped identify the weaknesses. Before official patches were released on October 21, 2026, a proof-of-concept video and early patched code had already become available.

Following the official patch, hundreds of purported proof-of-concept exploits appeared. Many reportedly did not function, creating additional work for defenders who needed to distinguish genuine threats from unreliable or AI-generated submissions.

This introduces an important secondary problem: AI can increase the volume of security information without necessarily improving its accuracy.

Security teams must now investigate both genuine exploit attempts and large quantities of potentially misleading vulnerability claims.


LLMJacking: How Cybercriminals Steal AI Computing Resources

The expansion of AI infrastructure has created another lucrative opportunity for financially motivated attackers.

LLMJacking involves compromising an organization's cloud accounts, AI credentials, or computing resources to obtain unauthorized access to large language models and other expensive services.

Instead of selling stolen personal information alone, attackers can monetize access to the computational resources that power modern AI applications.

Mandiant reported responding to multiple LLMJacking incidents during 2026. Some generated fraudulent charges exceeding $100,000 in a single day, while the most severe cases accumulated approximately $1 million in unauthorized resource consumption before remediation.

The attack lifecycle generally involves several stages.

Attack stage

Observed activity

Potential business impact

Initial access

Stolen credentials, session tokens, deceptive websites, exposed secrets

Unauthorized entry into cloud environments

Reconnaissance

Examination of repositories, cloud services, containers, and access logs

Discovery of valuable accounts and AI resources

Privilege escalation

Creation of service accounts or modification of identity permissions

Expanded access and persistence

Resource abuse

Activation of LLM services, GPU provisioning, quota increases

Unexpected computing expenses

Monetization

Deployment of API gateways, proxies, and relay infrastructure

Sale of unauthorized AI access

Attackers have used credentials obtained from public GitHub repositories and social engineering campaigns to access cloud environments.

After gaining entry, they may inspect available AI services, examine usage information, and identify opportunities to consume or resell computing resources.

In one incident, a threat actor used a strategy described as burn and rotate, repeatedly creating and deleting short-lived service accounts to circumvent usage restrictions.

Another observed vulnerability involved CVE-2026-40217, a sandbox escape in LiteLLM that could allow limited-access inference keys, known as Virtual Keys, to be elevated into administrative credentials.

The danger is that attackers do not necessarily need to install traditional malware to maintain access. Unauthorized identity permissions and service accounts can provide sufficient persistence.


LLM Mixers Create an Underground Market for Stolen AI Access

The monetization of LLMJacking has contributed to the emergence of underground services known as LLM Mixers.

These platforms sell unauthorized access to AI models through proxy infrastructure that connects customer requests to pools of compromised accounts and computing resources.

The business model resembles other specialized cybercrime marketplaces. Different operators can handle credential theft, initial access, infrastructure management, and customer fulfillment.

A mixer may distribute requests across multiple compromised environments, making it harder for defenders to identify abnormal activity associated with one account.

GTIG examined advertisements on Russian- and Chinese-language underground marketplaces and found that services offered access to multiple models, including frontier and jailbroken systems.


Reported prices ranged from $0.10 to $25 per million tokens, depending on the platform, model, and type of credentials.

Some vendors use subscription pricing instead of token-based billing. JetNight, for example, offers monthly access at different prices depending on the model. APIMart applies a fivefold price increase to input and output pricing when prompts exceed 100,000 tokens.

These services effectively transform stolen enterprise AI resources into a commercial product for other cybercriminals.

The broader implication is that AI infrastructure itself has become a target for financial exploitation. Organizations must therefore protect model credentials and cloud quotas with the same seriousness applied to financial accounts and production databases.


AI Is Transforming Information Warfare and Influence Operations

The threat extends beyond conventional cybercrime.

GTIG assesses with high confidence that state-sponsored groups, commercial Influence-as-a-Service providers, and inauthentic political monetization networks have moved toward routine operational use of LLMs.

AI-generated content can help operators overcome language barriers, create convincing synthetic personas, produce localized messaging, and distribute material across multiple platforms.

Three broad categories of activity are particularly important:

State-sponsored influence operations: Groups linked to Russia, China, and Iran use AI to support research, content generation, persona development, and coordinated distribution.

Commercial Influence-as-a-Service: Private contractors and marketing-style operations sell influence campaigns and automated platforms to clients.

Inauthentic political monetization: Profit-driven networks produce political content to maximize engagement, advertising impressions, click-through rates, and revenue rather than advance a particular ideology.


One platform investigated by GTIG in August 2026 used the Gemini API alongside other AI services to automate much of the influence campaign lifecycle. Its capabilities included persona creation, multimodal content generation, distribution, and interactive posting.

The system was designed to imitate grassroots participation by generating localized personas using regional language, slang, informal expressions, and deliberate spelling mistakes.

Another platform was designed to manage more than 100,000 accounts across major global social networks. Its architecture incorporated virtual smartphones, residential proxies, device-identity manipulation, disposable email addresses, and virtual phone numbers.

AI was also used to simulate ordinary user behavior, including scrolling, viewing content, and interacting with unrelated posts before accounts began distributing targeted messages.

The objective was to make automated profiles appear more authentic and reduce the likelihood of detection.

These findings demonstrate that the influence threat is not simply about producing more AI-generated articles or images. It involves building complete distribution systems capable of coordinating accounts, personas, and content at scale.

Importantly, higher content quality does not automatically translate into greater political influence. Audience trust, platform enforcement, distribution reach, and actual behavioral impact remain separate factors.


State-Sponsored Espionage Is Exploiting AI Development Workflows

AI is also creating new opportunities for supply chain attacks.

GTIG observed PRC-nexus groups using LLMs for reconnaissance, target research, exploit development, and analysis of stolen information. One contractor, LEAKY CASTLE, reportedly used AI to analyze and translate exfiltrated data, identify organizational relationships, and prepare intelligence reports for customers.

Another emerging technique involves abusing AI platforms and file-hosting infrastructure for staging malicious files, command-and-control communication, and data exfiltration.

DPRK-linked threat actors have also used generative AI to support fraudulent employment operations. These activities include producing resumes, preparing for interviews, and generating responses during recruitment screenings to impersonate Western IT workers.


A particularly concerning development involves the suspected MIDNIGHT NEPTUNE cluster and its targeting of cryptocurrency organizations in North America and Europe.

The attackers reportedly created fraudulent recruiter profiles using AI-generated images containing SynthID watermarks. Victims were directed toward deceptive meeting platforms that used ClickFix techniques to persuade them to execute diagnostic scripts.

After obtaining an initial foothold, the attackers accessed internal code repositories and modified .claude/settings.json configuration files. These changes were designed to trigger the download of the SOMBERMEME backdoor when users launched Claude Code.

The attackers also generated GitHub Personal Access Tokens using names associated with Microsoft Copilot to disguise lateral movement.

This illustrates an important evolution in enterprise security threats: development tools and AI coding assistants can become part of the attack surface.

Organizations must treat agent configuration files, repository permissions, extension systems, and automated execution hooks as security-sensitive infrastructure.


AI-Powered Phishing Lowers the Barrier for Cybercriminals

The commercialization of AI-assisted attacks is also changing phishing operations.

GTIG identified Jieyou, a Chinese-language Phishing-as-a-Service platform first observed in March 2026. Its June backend update introduced an LLM chatbot that allows customers to configure phishing panels and modify templates using natural-language instructions.

This removes some of the technical knowledge previously required to customize phishing infrastructure.

Instead of manually editing templates and scripts, affiliates can describe the changes they want and use the chatbot to generate modifications.

The platform's model-agnostic architecture integrates multiple frontier and open-source AI services through APIs.

This development reflects a broader trend in cybercrime: sophisticated capabilities are increasingly packaged into accessible commercial services.

As a result, defenders can no longer assume that a poorly skilled attacker will necessarily use technically unsophisticated infrastructure.


How Organizations Can Defend Against Agentic AI Cyberattacks

The growing use of autonomous offensive systems requires organizations to strengthen security across identities, software development, cloud infrastructure, and AI services.

Traditional perimeter security is insufficient when attackers can operate through legitimate credentials and trusted development tools.

A practical defense strategy should include the following measures:

Security priority

Recommended protection

Identity security

Enforce phishing-resistant multifactor authentication and monitor suspicious sessions

Cloud access

Apply least-privilege IAM policies and regularly audit service accounts

AI credentials

Store API keys in secure secret-management systems and rotate exposed credentials

Usage monitoring

Establish spending alerts, quota limits, and anomaly detection for AI services

Software security

Apply patches quickly and prioritize internet-facing systems

Development workflows

Protect agent configuration files, repositories, and automated execution permissions

Network defense

Monitor unusual outbound traffic, proxy usage, and data transfers

Incident response

Prepare procedures for disabling compromised accounts and stopping unauthorized inference

AI governance

Evaluate model access, data handling, third-party integrations, and agent permissions

Google describes a defense strategy combining model-level safeguards, threat intelligence, platform enforcement, and automated security operations.

Its Google AI Threat Defense architecture integrates Gemini models, Wiz risk prioritization, CodeMender remediation capabilities, and Mandiant intelligence. The approach uses different models according to the risk and computational requirements of individual security tasks.

Google also identifies Big Sleep and CodeMender as examples of AI-assisted vulnerability discovery and remediation. These initiatives illustrate how AI can support defensive security as well as offensive operations.

The company introduced its Secure AI Framework (SAIF) to provide a conceptual foundation for securing AI systems throughout their development and deployment.

For defenders, the objective should be to use automation to shorten detection and response times while preserving human oversight for sensitive decisions.


The Future of Cybersecurity: Defending Against Machine-Speed Attacks

The emerging agentic threat landscape is likely to become more complex as models improve their reasoning, tool-use capabilities, and ability to coordinate long-running tasks.

However, autonomous execution currently introduces weaknesses that defenders can exploit. GTIG notes that many offensive frameworks produce recognizable technical traces, unusual traffic patterns, and forensic artifacts. Their speed can come at the expense of stealth.

This creates an important distinction between financially motivated cybercriminals and espionage operators.

Cybercriminals may accept a higher probability of detection if automation allows them to reach more victims and generate greater financial returns. Intelligence-focused groups, by contrast, may remain more selective when operating inside sensitive environments where maintaining long-term access is more important than speed.

The balance may change as agent frameworks become more efficient and their execution patterns become harder to distinguish from legitimate enterprise automation.

The cybersecurity industry must therefore prepare for threats that combine human direction with increasingly autonomous execution.


AI Is Becoming Both a Cybersecurity Shield and an Attack Multiplier

Google's latest threat intelligence findings demonstrate that generative AI has become an operational component of modern cybercrime, espionage, vulnerability exploitation, and influence operations.

Multi-agent frameworks are helping attackers coordinate complex activities. Automated patch analysis is accelerating vulnerability exploitation. LLMJacking is turning stolen cloud resources into an underground AI economy, while LLM Mixers are commercializing unauthorized access to advanced models.


At the same time, AI-powered phishing and supply chain attacks are targeting the software and development environments organizations increasingly depend on.

The challenge for businesses is no longer simply preventing unauthorized access. It is securing the automated systems, identities, APIs, repositories, and computing infrastructure through which both legitimate and malicious agents operate.

For technology leaders, including Dr. Shahid Masood and the expert team at 1950.ai, these developments highlight a defining issue in the next phase of digital transformation: AI adoption and cybersecurity can no longer be treated as separate strategic priorities.

Key Takeaways

  • Threat actors are increasingly deploying autonomous multi-agent frameworks to coordinate cyber intrusions.

  • AI-assisted patch analysis is accelerating the discovery and exploitation of known vulnerabilities.

  • exploit_agent demonstrates how vulnerability research and exploitation can be connected into automated pipelines.

  • LLMJacking incidents have generated fraudulent charges exceeding $100,000 in a single day, with some cases reaching approximately $1 million.

  • Underground LLM Mixers are reselling unauthorized access to frontier AI models and compromised computing resources.

  • AI-generated personas and automated distribution systems are expanding the scale of influence operations.

  • Supply chain attacks increasingly target AI coding assistant configurations and development environments.

  • AI-assisted phishing services are making sophisticated campaign customization accessible to less technically experienced criminals.

  • Organizations need stronger identity protection, cloud monitoring, software patching, agent security, and automated incident response.

Future Outlook

The next generation of cyberattacks will increasingly combine human operators, autonomous agents, compromised infrastructure, and automated decision-making. As these systems become more capable, the time available to detect and contain an intrusion may continue to shrink.

Organizations that integrate AI-driven threat detection, secure development practices, continuous monitoring, and disciplined access controls will be better positioned to manage this changing environment.

The future of cybersecurity will depend not only on building more intelligent AI systems, but also on ensuring that their capabilities cannot be easily converted into scalable tools for exploitation.


Further Reading / External References

GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold

Google Finds Hackers Using AI Agents to Build and Run Exploits From Software Patches

Comments


bottom of page