It is three in the morning, and nobody is awake in the office. But somewhere on the internet, a piece of software is wide awake. It watches a vendor publish a security patch, reads the fix, and works out the flaw it was meant to close — the first step of an autonomous AI cyberattack.
Then it writes an attack for that flaw — and fires it off. No human typed a command. No human approved a step. That is what an autonomous AI cyberattack looks like, and Google’s threat researchers say they have already found one running in the wild.
The warning comes from the Google Threat Intelligence Group, known as GTIG, in a new AI Threat Tracker report produced with Mandiant and published on November 16. According to the report, hackers are already using large language models to shorten attack timelines, scale up their operations, and run multi-stage intrusions faster than ever before.
The message is blunt: the human is leaving the attack chain. And the clock, Google says, is now measured in months, not years.
The Exploit That Wrote Itself
The most striking evidence in the report is a single exposed server. On it, Google’s researchers found a framework they call exploit_agent — a multi-agent system built to find software flaws, build attacks for them, and launch those attacks, all on its own.
The system worked like a tireless researcher with no need for sleep. It watched widely used web software for new updates. When a patch appeared, it compared the code before and after the fix, spotted the vulnerability the patch was closing, drafted a working exploit, and ran it.
It was connected to both commercial and locally run large language models. At no point did a person approve each step. Google says this is the first time its researchers have seen threat actors use LLM frameworks to fully automate both vulnerability discovery and exploitation without a human in the loop.
Whoever ran the server remains unknown, and so does the scale of any intrusions it caused. But the system proves a point security teams have feared for years: the hardest part of hacking — turning a patch into a weapon — can now be automated.
Stage 5 of 7: Where Autonomous AI Cyberattacks Stand Today
Google sorts the rise of AI in cyberattacks into seven stages. The first four are history. It began with people simply prompting chatbots. Then AI was built into malware. Then it was woven into attack operations, with humans still in charge. Then came experiments with autonomous frameworks.
We are now at stage five, according to GTIG. At this stage, AI tools carry out several stages of an attack with what Google calls “minimal human-in-the-loop”. Multiple attack lifecycles are already being executed by LLM-powered tools.
Late 2025 was the turning point, when criminals first found real success in directly prompting large language models to identify vulnerabilities. In about a year, the field has moved from chatbots answering questions to autonomous systems engaging in breaches on their own.
Stage six is a forecast, not an observation. Google expects it within one to three months: simple, multi-agent-driven attacks with little or no human interaction. Stage seven follows in three to six months — sophisticated, fully autonomous attacks, arriving around the middle of 2027.
An Assembly Line for Breaking In
The report names the tools making this possible. Open-source, AI-driven penetration-testing frameworks such as ARTEX, CAIRN, CyberStrikeAI, PentAGI, and HexStrike AI are already being used in real intrusions, according to Google.
An orchestration agent runs the operation. It starts with reconnaissance, scanning external networks to find a way in. Once inside, the AI establishes persistence, quietly securing its foothold. Then it begins internal reconnaissance, mapping the network from within, and moves laterally from one system to the next.
ARTEX, the best known of these frameworks, has already drawn a response from the FBI. It was also named as a possible tool in the recent cyberattack on the European Commission, where hackers disrupted internal systems for more than ten days.
These frameworks are not theoretical. They are already being linked to real incidents, as wasl.tech reported this week. Google also says it has seen CAIRN used against targets in the United States aerospace sector, and a customised version of PentAGI used against organisations in Asia.
A Chinese-linked group, meanwhile, appears to be building its own automated attack tool, most likely for espionage.
Patch Day Has Become Race Day
Perhaps the most unsettling detail in the report is how attackers are turning the defence itself against defenders. Every time a vendor releases a patch, it quietly announces what was broken. AI agents now read that announcement in minutes.
Using a technique called differential analysis, the frameworks compare the patched software with the version before it. The difference between the two points straight at the flaw. The agent then builds an exploit for it — and can scale that attack across internet-facing servers without any human help.
The window between a patch and an exploit used to be measured in days or weeks. It is now measured in hours. And the volume is climbing: in a September analysis, Google found that reported vulnerabilities and exploits almost doubled between January and August 2026.
Patching quickly has always mattered. When the other side can weaponise a fix within hours, it becomes the single most important defence an organisation has.
Stealing AI to Power the Attacks
The report describes a second, stranger economy growing alongside the attacks. Financially motivated criminals are using a technique called LLMjacking: stealing AI model credentials and cloud resources from vulnerable enterprise environments, then reselling access to high-performance AI models through proxy services.
In other words, attackers are stealing other people’s AI to run their own AI attacks. The stolen compute becomes the engine for the next wave of intrusions, and the victim’s cloud bill pays for it.
State-sponsored actors are moving in a different direction. According to Google, they have begun routinely using large language models to generate synthetic media at scale, create localised personas, and automate the distribution of that material across platforms.
Why Defenders Still Have a Window
None of this means the defence is finished. Google’s researchers are careful to note what today’s AI attacks still get wrong. Current AI-enabled tools, they write, “trade stealth for velocity” — they move fast, but they generate activity that is easier to spot.
Autonomously executed attacks can also have a lower success rate than human-led ones. The danger is that the sheer reach and frequency of machine-run attacks compensates for that. A human hacker might try ten targets a week. An agent can try ten thousand.
The pattern is familiar from the wave of AI infrastructure breaches this autumn. The Fireworks AI breach exposed customer datasets and model weights, and the Railway breach saw stolen AI keys rack up $740,000 in fraudulent charges. In each case, the weakness was not the AI itself, but the systems and credentials around it.
The same lesson applies here. Microsoft’s own research into backdoored AI agents, which wasl.tech covered earlier this month, showed that even a handful of altered neurons can turn an agent rogue. The supply chain of AI — the models, the keys, the frameworks — is now part of the attack surface.
For defenders, the practical advice has not changed, but the urgency has. Patch fast, because the gap between a fix and an exploit is now hours. Watch for machine-speed activity, because the noise of a fast, clumsy agent is still detectable. And treat AI credentials like the keys they are.
What Happens Next
Google’s timeline gives the industry a rough countdown. Within one to three months, expect the first wave of simple attacks with little or no human involvement. Within three to six months, expect those attacks to become more sophisticated — and harder to detect.
That puts the turning point somewhere in the first half of 2027. By then, the question will no longer be whether AI agents can attack on their own. It will be how many are doing it, and how quickly defenders can spot them.
There is still time to prepare. The tools attackers use today are fast but noisy, and the fully autonomous attacks Google predicts have not yet become the norm. The report is a forecast, not a surrender — and forecasts are useful only if someone acts on them.
But the direction of travel is clear. The hacker of the future may not be a person at all. It may be a system that never sleeps, never hesitates, and never asks permission. According to Google, it is already knocking on the door.



