According to GTIG, attackers are now fully automating vulnerability discovery and exploitation for the first time. Google expects sophisticated autonomous AI attacks to emerge within six months.
Attackers are increasingly using agentic workflows built on AI frameworks to automatically identify and exploit security vulnerabilities. The Google Threat Intelligence Group (GTIG) reports this trend in its latest AI Threat Tracker. While AI was initially used in cyberattacks mainly on an experimental basis, Google is now observing widespread adoption by both financially motivated and state-sponsored threat groups. Within the next three months, attackers are likely to be capable of launching basic attacks that require little to no human intervention and rely primarily on multiple AI agents. Google expects sophisticated, fully autonomous AI attacks to emerge within the next six months.
“In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”
exploit_agent Analyzes Software Updates
GTIG cites a multi-agent system called exploit_agent as one example. It integrates several leading language models, including locally hosted ones. Google discovered the system on an unsecured staging server belonging to the attackers. The server also contained stolen victim data, reconnaissance findings, web shells, and password-cracking scripts.
The system analyzes updates to widely used web applications for signs of security vulnerabilities and attempts to exploit them automatically. It was used against websites in the media, entertainment, and education sectors in France.
Google also reports that attackers are systematically abusing open-source penetration testing tools such as ARTEX, CAIRN, CyberStrikeAI, and HexStrike AI to automate attacks against internet-facing systems with easily exploitable vulnerabilities. CAIRN, for example, was used in an attack targeting the U.S. aerospace industry. Last week, the FBI warned that multiple threat actors had been using ARTEX since August to target organizations in the financial services, government, technology, defense, and healthcare sectors.
Less Time to Patch Vulnerabilities
According to Google, the use of these frameworks significantly reduces the time defenders have to patch vulnerabilities. At the same time, it creates false leads. Following the disclosure of the PHP2Shell vulnerabilities, hundreds of purported proof-of-concept exploits surfaced, most of them AI-generated. Defenders had to spend time assessing their validity.
GTIG also observed attacks by the group UNC6836 against cryptocurrency companies in North America and Europe. The group is suspected of having links to North Korea. After gaining access to their targets, the attackers modified the settings of the Claude Code programming tool so that it would load the SOMBERMEME backdoor on startup.
AI is also being used routinely in influence operations. Google further reports that LLMjacking, in which attackers use stolen credentials to access AI services at their victims’ expense, has become widespread.
(Editorial Team)