Google Reveals How Hackers Are Using AI Agents to Automate Real-World Cyberattacks

Blog WriterCybersecurity News - Original News Source is cybersecuritynews.com


Hackers are no longer just asking AI chatbots for help. They now run AI agents that can scan, break in, and move through networks with very little human input.

In its latest AI Threat Tracker, published on November 16, 2026, the Google Threat Intelligence Group (GTIG) and Mandiant said AI misuse has reached a “critical inflection point” across cybercrime, spying, and influence campaigns.

The findings come from Mandiant incident response cases, global tracking of threat groups, and Google’s own platform defenses. Google says attackers are using AI to shorten attack timelines, grow their operations, and run multi-stage intrusions at machine speed.

LLM maturity timeline (Image Source: Google)

Google AI Agent Cyberattacks

Google says threat actors are actively using open-source, multi-agent hacking frameworks such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI. GTIG saw CAIRN and PentAGI used in the wild and spotted testing of CyberStrikeAI and HexStrike AI.

Tool Main purpose Reported use by threat actors Observation status
ARTEX Multi-agent penetration testing that turns an operator’s goals into commands Automated network discovery, privilege escalation, lateral movement, and data theft tasks Used in real-world intrusions
CAIRN Coordinates specialized agents through a shared workspace and an Observe, Orient, Decide, Act loop Navigated attack paths against targets, including organizations in the US aerospace sector Used in real-world intrusions
PentAGI Automates penetration testing through agents that divide objectives into smaller tasks A customized version performed reconnaissance and vulnerability scanning against organizations across Asia between late August and early October 2026 Deployed in the wild
CyberStrikeAI Connects AI models to more than 100 security tools, attack recipes, and custom scripts Tested for coordinating and automating attack workflows Threat actor experimentation observed
HexStrike AI Uses Model Context Protocol to connect AI clients to more than 150 offensive security tools Tested for AI-driven reconnaissance, web testing, exploitation, and binary analysis Threat actor experimentation observed
Agent Zero General-purpose agent framework supporting autonomous code execution and dynamic tool creation Explored as part of threat actor interest in autonomous attack systems Threat actor experimentation observed

CAIRN was used in intrusions against US aerospace targets. A custom platform built on PentAGI scanned many organizations in Asia between late August and early October 2026.

ARTEX remains the best-known example. Cyber Security News earlier reported how hackers use ARTEX AI agents to automate Active Directory attacks, with single bursts of activity lasting just 2 to 25 seconds.

Wiz researchers also tracked an ARTEX-using actor, overlapping with Microsoft’s Storm-0940, that went after Okta, Entra, and Cognito identity systems at IT and software vendors.

There is a weak spot. Google says these tools trade stealth for speed, leaving noisy traffic, known names, and clear code traces behind. Even so, they still reach their goals, and Google expects less human involvement and better stealth over time.

Multi-agent LLM offensive frameworks
Multi-agent LLM offensive frameworks (Image Source: Google)

For the first time, GTIG found attackers using AI to find and exploit flaws with no human in the loop. An exposed staging server hosted a framework called “exploit_agent” that compares web software code before and after patches, finds bugs, writes exploits, and runs them on its own.

Russian comments in its setup files pointed to a pilot project. The operator focused on French media, entertainment, and education websites.

Real cases back this up. After two PaperCut flaws (CVE-2026-81578 and CVE-2026-82078) were exploited as zero-days in August 2026, GreyNoise reported an AI-assisted campaign that hit hundreds of organizations and reached domain admin within minutes.

Zero-day exploits climbed from 8 per month in 2025 to 11 per month in 2026, then jumped to 22 in August.

Money-driven groups keep stealing AI keys and cloud access, then reselling it through “LLM Mixers.” Google’s LLMJacking research found fake charges topping $100,000 in a single day, with some victims reaching $1 million. Our earlier report on LLM Hijacking attacks and million-dollar cloud bills explains how attackers test stolen keys with tiny requests to avoid alerts.

A suspected North Korean cluster, MIDNIGHT NEPTUNE, used fake recruiter profiles with AI-made images and ClickFix lures to break into crypto firms. It then changed .claude/settings.json files so Claude Code would download the SOMBERMEME backdoor at startup. It also hid its tracks with GitHub tokens named after Microsoft Copilot.

Influence groups now use AI as part of their normal work. One contractor used the Gemini API while building a platform designed to run over 100,000 social media accounts with cloud phone farms. Fake accounts were “aged” by AI that copied normal scrolling and viewing habits before posting political messages.

Google SecOps has released behavior-based rules to detect ARTEX. Mandiant’s five-layer defense guide advises teams to lock down admin systems, use short-lived credentials, limit access, split networks, and watch for sudden spikes in API and identity activity.

For AI services, teams should set budget alerts, scan code for leaked keys, and treat any unexpected AI service activation as a serious warning.

As Google notes, the methods have not changed. AI simply lets attackers chain them together much faster, so defenders have less time to respond.

Give your SOC the alert context to stops incidents quickly. Integrate Threat Intel from 16K SOCs.

The post Google Reveals How Hackers Are Using AI Agents to Automate Real-World Cyberattacks appeared first on Cyber Security News.