It was an ordinary Tuesday in a French newsroom. A software vendor published a security patch — the kind of routine update that usually passes without a second thought.
Somewhere else, a machine was watching. It read the patch, worked out exactly which flaw the update had just closed, wrote an exploit for it, and fired it off.
No human typed a command. No one was even awake to watch it happen. According to Google, autonomous AI cyberattacks are no longer a warning about the future — they are a case file from the present.
What Is exploit_agent? Inside the Autonomous AI Cyberattack Framework
The discovery sits at the centre of Google’s fourth-quarter AI Threat Tracker, published on November 16 by the Google Threat Intelligence Group, known as GTIG, together with Mandiant.
Researchers found a multi-agent framework, which they have dubbed exploit_agent, running on an exposed staging server. It connects to several large language models — both frontier systems and models hosted locally — and it works on its own.
Its method is simple, and that is what makes it dangerous. The framework watches popular web software and compares the code before and after a vendor announces a patch.
The difference between the two versions tells it what was fixed. From that difference it drafts an exploit, then runs it — all without a person in the loop.
The staging server held more than the framework itself. Google found data taken from victims, along with scripts for internal reconnaissance, password brute-forcing, installing webshells, and even removing other attackers’ webshells.
Configuration comments in the framework were written in Russian, and the victim telemetry pointed at French websites in media, entertainment and education. Google did not name a group or a country behind the operation.
Autonomous AI Cyberattacks Target Patched Flaws, Not Zero-Days
The report corrects a common assumption about where the danger lies. The automation is not mainly hunting for zero-days — flaws nobody knows about yet.
It is aimed at n-days: flaws that already have a patch. A patch tells an attacker exactly where to look, because it shows precisely which lines of code changed.
That makes the gap between a vendor shipping a fix and a customer installing it the entire game. If your patch cycle is measured in weeks, that gap is now the whole attack surface.
Two examples in the report show how fast the window is closing. Attackers exploited two PaperCut print-server flaws, CVE-2026-81578 and CVE-2026-82078, as zero-days in late August.
Researchers and the security firm Huntress built working exploits within a day of disclosure, according to the report. Four days later, GreyNoise recorded a campaign against hundreds of organisations, with the attacker moving from a first foothold to domain administrator in minutes.
Two PHP flaws, CVE-2026-109618 and CVE-2026-109619, cut the other way. They were found with LLM-assisted analysis, and before official patches landed on October 21, a proof-of-concept video and early patched code were already public.
Hundreds of proof-of-concept exploits published in the following days were claimed but did not work — most likely generated by AI — which handed defenders verification work instead of relief.
The Numbers Behind the Autonomous AI Cyberattack Surge
GTIG’s numbers show a threat that is growing fast. Overall vulnerability disclosures and exploits nearly doubled between January and August 2026.
Zero-day exploitation rose more slowly at first — from a 2025 average of eight a month to eleven a month in 2026 — before jumping to 22 in August alone.
The tooling is spreading beyond a single framework. GTIG observed CAIRN used against US aerospace targets, and a customised PentAGI platform scanning organisations across Asia between late August and early October.
Researchers also saw experimentation with CyberStrikeAI and HexStrike AI, and evidence that a China-linked group is building its own automated framework for espionage.
A separate finding in the report concerns a suspected North Korean cluster, MIDNIGHT NEPTUNE, which changed Claude settings files inside internal code repositories at cryptocurrency organisations — a reminder that AI coding tools themselves are now part of the attack surface.
The FBI and Amazon Sound the Same Alarm on Autonomous AI Cyberattacks
Google was not alone. The FBI and Amazon Threat Intelligence published parallel warnings on November 16, and the picture they paint is the same.
The FBI had already warned on November 13 about ARTEX, an AI penetration-testing framework turned into a weapon. Its alert said ARTEX intrusions reached privileged access in under three hours.
The same alert linked ARTEX to attacks that affected roughly 67,000 people at South Korean financial institutions in August, and to a ConnectWise supply-chain attack that compromised more than 2,250 downstream Microsoft tenants. We covered the FBI ARTEX warning in detail when it landed.
Amazon’s honeypots caught two ARTEX campaigns between October 3 and November 5. One harvested AI credentials at industrial scale. The other built real-time deepfakes to beat identity checks at two crypto exchanges.
The deepfakes landed a face within seven pixels of the target and ran at 11.5 frames per second on a single consumer graphics card — though the exchanges’ server-side checks rejected every attempt.
LLMJacking: The Profitable Side of Autonomous AI Cyberattacks
There is also money in it. Mandiant tracked LLMJacking incidents — where criminals break into a company’s cloud environment and spend its AI quota — with single-day fraudulent charges above $100,000.
Cumulative abuse reached $1 million before it was stopped in some cases. The stolen compute is believed to power further AI work, including model training and distillation.
It is the same pattern Google has seen before. The Fireworks AI breach showed attackers going after AI provider credentials — proof that AI infrastructure itself has become a target worth stealing.
Why Autonomous AI Cyberattacks Are Loud — and What Stops Them
For all their speed, these frameworks have a weakness. GTIG notes they currently trade stealth for velocity: unrestricted autonomous execution generates noisy forensic traffic, relies on recognisable identifiers and leaves code-level traces.
That makes them easier to detect — for now. But Google expects the frameworks to get quieter as they mature, which is why the report frames the current stage as a threshold rather than a peak.
Microsoft’s own research points the same way: its Lazarus backdoor study showed how few changes it takes to turn an AI agent rogue, and how hard that can be to spot.
The defence that held in every case Google and the FBI described is unglamorous. Patch fast. Keep credentials out of code. Close unauthenticated interfaces and inference endpoints.
As Amazon CISO CJ Moses put it, the agents succeeded where basic security hygiene was missing — default or hardcoded credentials, keys left in code and public artifacts, unauthenticated interfaces, and unpatched software.
Google has also introduced Google AI Threat Defense, a security system designed to detect and address threats using AI — fighting machine-speed attacks with machine-speed defence. The full report is covered by Techaeris.
What Happens Next: Fully Autonomous Attacks Within Months
The report’s timeline is the part that should worry defenders most. A chart in it places current activity at Stage 5: multiple attack stages conducted by LLM-driven tooling in the wild, with minimal human involvement.
Stage 6 — multi-agent attacks with little to no human interaction — is projected within one to three months. Stage 7 — sophisticated, fully autonomous attacks — is projected within three to six months.
Those later stages are forecasts, not observations. But the direction is clear: the human is leaving the loop, and the window for patching is shrinking with every step.
The question for every organisation is no longer whether AI will be used against it. It is whether its patching and hygiene can keep up with a machine that never sleeps, never waits for a maintenance window, and reads every patch the moment it lands.


