Skip to content
Google catches AI agents hacking with no human in the loop

Google catches AI agents hacking with no human in the loop

E27.Co • November 18, 2026

GTIG's latest AI Threat Tracker finds exploit pipelines that diff patches, write exploits and fire them unaided, shrinking defenders' response time

The hacker hunched over a keyboard is becoming a supervisor. According to Google, more and more of the actual break-in work is now being done by software agents that scan, exploit and adjust on their own.

In its latest AI Threat Tracker, published early this week, Google Threat Intelligence Group (GTIG) and its incident response arm Mandiant say the adversarial use of AI has reached a “critical inflection point”. Attackers have moved past asking chatbots for help with phishing emails. They are now deploying autonomous, multi-agent frameworks that chain together entire intrusions. For the first time, GTIG says, it has seen threat actors fully automate vulnerability discovery and exploitation with no human in the loop.

Also Read: The AI agent boom is exposing Southeast Asia’s startup codebase problem

Southeast Asia is not a bystander in this story. The report names a campaign that swept organisations across Asia, stolen AI compute in the Asia Pacific, a Chinese-language phishing kit built for e-commerce fraud and North Korean operators going after crypto firms. Those are four of the region’s softest spots, all on one page.

Penetration-testing tools turned into attack crews

The tools at the centre of the report were built for the good guys. ARTEX, CAIRN, CyberStrikeAI and PentAGI are open-source “agentic pentesting” frameworks, designed to let security teams simulate attacks with teams of AI agents that divide up the work. GTIG says it has seen CAIRN and PentAGI deployed in real intrusions, and attackers experimenting with CyberStrikeAI and HexStrike AI, a harness that uses the Model Context Protocol (MCP) to plug an AI model straight into more than 150 offensive security tools.

The shared recipe is simple and alarming. The agents find low-complexity vulnerabilities, generate single-use payloads on demand and adjust their tactics from real-time feedback. GTIG observed CAIRN in intrusions against US aerospace targets. The same framework recently featured in a NorForge investigation that put the cost of a break-in at roughly US$50 in AI tokens.

Asia was squarely in the frame. Between late August and early October, GTIG tracked an actor running a customised version of PentAGI to carry out systematic reconnaissance and vulnerability scanning against a wide range of organisations in Asia. The report also flags a group with links to China building its own automated attack framework, likely for espionage.

There is one consolation, for now. GTIG notes that these frameworks “trade stealth for velocity”: they generate noisy traffic and leave recognisable traces that defenders can catch. Google expects that to change as the tools and models mature. In other words, the window in which these attacks are easy to spot is closing.

The patch window is shrinking from weeks to days

The most striking find was sitting on an exposed staging server. GTIG discovered a framework its operators called exploit_agent, wired to several frontier and local AI models. It continuously compares popular web software before and after each patch, works out what was fixed, drafts an exploit and fires it, all without human intervention. Configuration notes were written in Russian, and victim data pointed to French media, entertainment and education websites. The kit even included scripts to remove webshells planted by rival attackers.

The statistics show the effect. GTIG found that overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits, flaws attacked before a fix exists, rose more slowly, from an average of eight a month in 2025 to 11 a month in 2026, before jumping to 22 in August alone.

Also Read: Equal spend: why AI budgets need a security twin

Extracted Entities

Attack Types (1)

Companies (1)

Countries (1)

Industries (2)

MITRE ATT&CK (1)

Platforms (1)