Google and Amazon Uncover Hackers Using AI Agents to Automate Cyberattacks

Google and Amazon have independently uncovered cybercriminals using artificial intelligence agents to automate cyberattacks, steal access credentials, exploit software vulnerabilities, and attempt fraud at cryptocurrency exchanges.

The findings, detailed in separate Google and Amazon security reports published November 16 and 13, show attackers moving beyond basic AI chatbots toward systems capable of performing complex hacking tasks with limited human supervision.

Both investigations identified ARTEX, an open-source AI-powered penetration-testing framework that connects AI agents with security tools to automate vulnerability discovery and other operations. Google also documented attackers using similar frameworks, including CAIRN and PentAGI.

AI Agents Used to Steal Credentials and Exploit Vulnerabilities

Amazon detected suspicious activity between October 3 and November 5, 2026, through MadPot, its global network of decoy systems designed to monitor cyber threats.

Records captured from AI agents described campaigns focused on stealing credentials used to access commercial AI models and cloud services. The attackers sought to reuse stolen keys to power additional AI-driven attacks without paying for the computing resources themselves.

The agents searched exposed applications, tested credentials, and evaluated what compromised accounts could access. Amazon assessed with moderate confidence that the campaigns involved ARTEX and potentially separate Chinese-linked threat groups.

However, Amazon cautioned that the records contained the agents’ own descriptions of their activities, meaning not every reported compromise could be independently verified.

Google’s investigation provided additional evidence of autonomous attacks operating in real-world environments.

Researchers observed CAIRN being used in intrusions targeting organizations including the U.S. aerospace sector, while a customized PentAGI-based platform conducted vulnerability scanning against organizations across Asia.

Google also discovered a system called exploit_agent that automatically compared patched and vulnerable software, identified weaknesses, generated exploits, and executed attacks without human intervention.

The company said this was its first observation of threat actors fully automating vulnerability discovery and exploitation using AI frameworks.

Google additionally highlighted a PaperCut software attack in which GreyNoise researchers identified at least 440 compromised installations across 48 countries. During one phase, the attacker compromised 11 organizations in just 26 seconds.

Stolen Cryptocurrency & AI Resources

Amazon also uncovered a separate ARTEX campaign targeting identity verification systems at two cryptocurrency exchanges.

The attackers used AI-generated videos and altered identity documents in attempts to impersonate legitimate users and bypass security checks. Some attempts involved manipulating facial movements to fool systems that verify whether a real person is present.

Amazon reported that one verification provider rejected every recorded attempt to bypass its facial verification system.

Meanwhile, Google documented a growing criminal market involving stolen AI computing resources, commonly known as LLMJacking.

Mandiant investigated incidents in which attackers compromised cloud environments to access commercial AI models and computing infrastructure without authorization. Some victims accumulated more than $100,000 in fraudulent charges within a single day, while cumulative resource abuse reached $1 million in the most severe cases.

Hackers were also reselling access to compromised AI services through underground platforms known as LLM Mixers.

Google identified additional AI-related threats, including suspected North Korean hackers manipulating Anthropic’s Claude Code configuration files to install malware and commercial operators developing AI-powered political influence platforms. One proposed system was designed to manage more than 100,000 social media accounts.

Despite the growing capabilities of autonomous hacking tools, Google warned that they remain imperfect. AI-generated code can fail, models can produce inaccurate results, and automated attacks often leave detectable evidence.

Both companies emphasized that attackers continue exploiting familiar security weaknesses, including exposed credentials, poorly configured applications, and unpatched software. What AI agents are changing is the speed and scale at which those weaknesses can be found and exploited.