Cybercriminals are increasingly using autonomous artificial intelligence agents to speed up cyberattacks and reduce the amount of manual work required from human operators.
Google Threat Intelligence Group (GTIG) and Mandiant observed a suspected financially motivated threat actor compromise an organization’s cloud infrastructure and deploy a multi-agent attack framework.
The attackers were able to plan, build, and execute a large credential-harvesting campaign in less than six hours.
The campaign shows how threat actors are moving beyond basic AI prompts and beginning to use agentic AI systems that can independently perform several parts of an attack.
Google said these systems can manage scanning operations, solve technical problems, make decisions, and continue working with limited human involvement.
AI Agents Mass-Steal Credentials
According to the report, the attackers used an AI coding chatbot together with prompts and preconfigured instructions written in Markdown files.
These instructions acted like operational playbooks that told the AI agents how to perform different stages of the campaign.
The autonomous framework scanned internet-facing systems for weaknesses and automatically attempted to harvest credentials.
The agents were also capable of managing vulnerability-scanning pipelines, troubleshooting problems in real time, and rotating IP addresses without requiring an attacker to manually control every step.
By operating from previously compromised cloud infrastructure, the attackers were also able to send malicious traffic through legitimate IP addresses.
This can make malicious activity more difficult for defenders to identify because the traffic may appear to originate from trusted cloud environments.
The operation ultimately compromised thousands of third-party credentials, according to Mandiant.
GTIG separately discovered an exposed command-and-control server hosting another automated reconnaissance and credential-management framework called “Recon.”
The system contained specialized agent configuration files and later developed into a working dashboard capable of organizing, validating, and managing more than 23,800 harvested secrets, including API keys for cloud and AI services.
The growing use of AI agents represents an important change in cyberattack operations. Traditional automated tools normally follow predefined scripts.
Agentic AI systems can instead analyze situations, make decisions, troubleshoot failures, and adjust their actions while an operation is running.
GTIG said adversaries are increasingly creating multi-agent frameworks that can autonomously manage scanning pipelines and conduct credential theft at scale.
AI is also being used across other attack stages, including reconnaissance, social engineering, malware development, vulnerability research, and post-exploitation troubleshooting.
This increased automation can significantly shorten the time available for security teams to detect and stop attacks.
A campaign that previously required attackers to manually configure tools and respond to errors can now move through several stages with much less human involvement.
Google said it disrupted some of the observed malicious activity by disabling associated assets and strengthening protections designed to prevent its AI systems from assisting similar attacks.
For organizations, the findings highlight the need for stronger monitoring of cloud infrastructure, exposed credentials, API keys, unusual automated scanning, and abnormal activity coming from trusted cloud services as AI-driven cyberattacks continue to become faster and more automated.
Detect 58% more threats with fresh intelligence from 16K+ organizations. Integrate TI Feeds in you SOC
