Google Says Cyberattackers Are Moving to Agentic AI as One Campaign Harvested Credentials in Under Six Hours
Google Threat Intelligence Group says cybercriminal and nation-state actors are moving beyond basic LLM prompting toward agentic and automated attack workflows, including a campaign that moved from compromised cloud infrastructure to mass credential harvesting in less than six hours.
Why it matters: Google is documenting a concrete transition from criminals using LLMs as assistants toward adversaries building autonomous and multi-agent cyber workflows.
Cyberattackers are beginning to move from simply asking artificial-intelligence models for help toward building agentic systems that automate parts of real-world attack operations, according to new threat intelligence from Google.
Google Threat Intelligence Group says its observations during the second quarter of 2026 show a shift from basic LLM prompting toward agentic AI workflows and AI-enabled automation. The distinction matters because agentic systems can coordinate tasks, react to errors, manage workflows and continue executing with less human intervention.
In one case investigated by Mandiant, a financially motivated attacker compromised cloud infrastructure and used AI-assisted automation to build and execute a mass credential-harvesting campaign in less than six hours. Google separately identified an automated reconnaissance and credential-management framework that had organized more than 23,800 harvested secrets, including cloud and AI-service API keys.
Google also says software supply chains and AI systems themselves are becoming targets, with attackers pursuing proprietary models, research, prompts, source code and other AI intellectual property. The company reports increasingly large attempts to replicate proprietary model capabilities through automated querying.
There is an important limit to the findings: GTIG says it has not yet observed attackers deploying fully autonomous zero-day exploitation pipelines against real-world targets. The evidence instead points to human attackers inserting AI into more stages of existing operations.
For defenders, the practical consequence is compressed response time. If attackers can automate hours or days of manual work, security teams have less time to detect and contain an intrusion before it advances.