Google's threat researchers say hackers have crossed a line the security industry has been warning about for years: AI agents are now finding software flaws, writing exploits for them and launching attacks with no human involved at any step.
The finding comes from the Google Threat Intelligence Group (GTIG) and Mandiant's latest AI Threat Tracker, published on November 16 and titled "Agentic Threat Activity Crosses a Significant Threshold." In it, the researchers say that in the past quarter they saw threat actors "using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time."
That is a sharp change from just two months ago. In the previous edition of the quarterly tracker, published in September, GTIG said it had not observed fully autonomous attack pipelines used against real targets. This quarter, it says, that changed.
The framework that reads patches and writes exploits
The headline example is a framework Google calls exploit_agent, found running on an exposed staging server wired into several frontier and locally hosted large language models. According to GTIG, it keeps comparing popular web apps' code before and after each patch announcement, analyses the differences to identify the likely flaw, drafts an exploit and then runs it automatically.
Material on the server suggested its operator was interested in French websites in the media, entertainment and education sectors. Google did not name the operator or give a success rate for the framework. One configuration comment was written in Russian.
The same technique is spreading. Google says researchers used language models to build working exploits for two PaperCut print-management flaws (CVE-2026-81578 and CVE-2026-82078) within a day of their disclosure in late August, after the flaws were exploited as zero-days.
Machine-speed intrusions are now routine
GTIG says it has observed open-source, multi-agent offensive frameworks being used in live intrusions, including CAIRN, used in an attack against organisations in the US aerospace sector, and PentAGI, used by a threat actor for reconnaissance and vulnerability scanning against organisations across Asia between late August and early October. It also saw experimentation with CyberStrikeAI and HexStrike AI.
The numbers are stark. Zero-day flaws exploited in the wild rose from an average of eight a month in 2025 to eleven a month in 2026, reaching 22 in August, while overall vulnerability disclosures and exploits nearly doubled between January and August 2026. The report also describes LLMJacking, in which attackers hijack cloud accounts to run up charges on AI services, with fake bills topping $100,000 in a single day and some victims reaching $1 million.
What comes next
Google's projection is blunt: within three months, threat actors will be able to launch unsophisticated attacks requiring "little to no human interaction," and sophisticated, fully autonomous agentic attacks are likely within six months. The FBI warned last week that multiple threat actors have been weaponising ARTEX, another agentic penetration-testing framework, since August, in attacks spanning finance, government, technology, defence and healthcare.
The automation has a weakness, the researchers note. Tools built for speed trade away stealth, leaving noisy traffic, recognisable tool names and clear code traces behind. Google has released behaviour-based detection rules for ARTEX, and Mandiant advises teams to lock down admin systems, use short-lived credentials, limit access, split networks and watch for sudden spikes in API and identity activity.
The practical message for everyone else is simpler: install security updates as soon as they appear. The gap between a patch being published and an exploit being written for it is now measured in hours — and increasingly, no human is doing the writing.



