Az | EN

Google Warns: AI Agents Could Soon Take Control of Cyberattacks

Cəmil Hüseynzadə
19 November 2026 12:17
108 views
Google Warns: AI Agents Could Soon Take Control of Cyberattacks

Cybercriminals are increasingly using agent-based workflows powered by AI attack infrastructure to automate the discovery and exploitation of security vulnerabilities. This allows attacks to be carried out at machine speed.

Although AI was initially used experimentally in cyberattack operations, Google’s Threat Intelligence Group (GTIG) has observed both financially motivated and state-sponsored threat groups adopting the technology on a wider scale.

Notable examples include the use of CAIRN, an open-source, multi-agent penetration-testing framework, in an attack targeting the US aerospace sector, and suspected North Korean hackers using Claude Code to distribute malware.

Researchers predict that within the next three months, threat actors could launch relatively simple attacks relying heavily on multiple AI agents and requiring little or no human interaction. They also expect complex, fully autonomous agent-based attacks to emerge within the next six months.

In a warning issued last week, the FBI reported that multiple threat actors had begun weaponizing the ARTEX AI penetration-testing framework in August. The attacks targeted organizations across finance, government, technology, defense, healthcare and other sectors.

Using AI to Find and Exploit Security Vulnerabilities

Google says attackers are systematically abusing open-source penetration-testing frameworks, including ARTEX, CAIRN, CyberStrikeAI and HexStrike AI, to carry out fully automated attacks.

These frameworks have been used to launch adaptive attacks against internet-facing endpoints vulnerable to relatively low-complexity flaws that large language models (LLMs) can easily exploit by generating attack code.

One example highlighted in GTIG’s AI Threat Tracker report involved hackers using a multi-agent system called “exploit_agent” to fully automate vulnerability discovery and exploitation in attacks targeting media, entertainment and education websites in France.

Google said GTIG identified, for the first time last quarter, threat actors using LLM frameworks to automate vulnerability discovery and exploitation without human intervention.

More specifically, the researchers discovered an internet-facing staging server hosting “exploit_agent”, a multi-agent vulnerability research and weaponization framework integrated with several leading and locally hosted LLMs.

Google found the infrastructure on an attacker-controlled, internet-facing staging server. The server also contained data stolen from victims, reconnaissance information about targets, web shell tools and brute-force attack scripts.

The system analyzes software updates for popular web applications to identify signs of potential vulnerabilities and automatically attempts to exploit them.

Google warns that the use of such frameworks in cyberattacks significantly reduces the time available for security teams to patch their systems. It also creates additional challenges by generating false leads.

Researchers pointed to the recent PHP2Shell case as an example. Shortly after PHP disclosed two vulnerabilities, hundreds of purported proof-of-concept (PoC) exploit codes emerged, most of them AI-generated. Security teams had to verify these codes and develop protective measures, placing a significant additional burden on defenders.

Other Trends in AI Usage

Beyond automated attacks, Google identified several attacks by UNC6836, a suspected activity cluster linked to “Midnight Neptune”, targeting cryptocurrency companies in North America and Europe using a modified version of Claude Code.

After gaining initial access, the attackers changed Claude Code’s settings so that launching the application would trigger the download of a backdoor known as SOMBERMEME.

Google also reports that state-sponsored groups, commercial vendors and political networks are now routinely using AI in influence operations.

Another growing concern identified by researchers is LLMjacking, the unauthorized use of large language model resources. Attackers are increasingly exploiting this activity as a source of revenue, imposing significant costs on the affected organizations.

© copyright 2022 | tech.az | info@tech.az