Google Says AI Agents Just Carried Out the First Fully Autonomous Cyberattack
Key takeaway: Google’s threat intelligence team says attackers have run the first fully autonomous cyberattacks it has ever observed. A multi-agent AI framework called exploit_agent reads security patches, works out what flaw each patch fixed, writes an exploit, and fires it — with no human in the loop. The finding, published November 16, 2026, shrinks the safe window after a patch from weeks to hours.
What Google found
On November 16, 2026, the Google Threat Intelligence Group (GTIG) and Mandiant published their quarterly AI Threat Tracker, titled “Agentic Threat Activity Crosses a Significant Threshold.” Its headline finding is a first: “In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”
The evidence is a framework Google calls exploit_agent, found running on an exposed staging server its operator left reachable on the internet. It is a multi-agent vulnerability research and weaponization system that connects to several frontier and locally hosted large language models. The same server held data stolen from victims, plus scripts for what comes after a break-in: internal reconnaissance, brute-forcing, installing webshells, and removing other attackers’ webshells.
Victim telemetry on the server suggests the operator was specifically interested in French websites in media, entertainment, and education. Google has not named the operator or the country behind it. The only clue to origin sits in a configuration file, where comments in Russian ask for the language models’ API keys to be entered for “auto-exploitation (exploit_agent pilot).” The Register describes the operator as suspected Russian-speaking.
The finding lands a day after California sued OpenAI over a separate case of AI agents acting without human control — read our coverage of California’s lawsuit over the Hugging Face hack.
How exploit_agent turns a patch into an attack
Exploit_agent runs what Google calls continuous differential analysis of popular web application software. The loop has four steps, and no person approves any of them:
| Step | What the framework does |
|---|---|
| 1. Watch | Monitors popular web software for new security patches |
| 2. Compare | Diffs the code before and after each patch announcement |
| 3. Reason | Works out which vulnerability the patch was fixing |
| 4. Execute | Drafts an exploit for that flaw and runs it automatically |
The example Google chose to print shows the framework inspecting access-control code in Spectra (Ultimate Addons for Gutenberg), a WordPress plugin at version 2.19.25 with more than a million active installations. Google does not say a flaw was found in that plugin or that it was exploited — the screenshot shows the agent reading the code, not breaking in.
How this differs from earlier AI-assisted hacking
Attackers have used AI for years — to write phishing emails, generate malware snippets, and speed up reconnaissance. In all of those cases a human still picked the target, checked the output, and pressed the button. Exploit_agent closes that loop: the agents find the flaw, build the weapon, and fire it.
It is not the only such tool. GTIG says threat actors are now abusing several open-source multi-agent frameworks in live intrusions: ARTEX, CAIRN, CyberStrikeAI, and PentAGI, with experimentation underway on HexStrike AI. GTIG observed CAIRN used in intrusion activity against targets including the aerospace sector in the US, and tracked a customized PentAGI-based platform conducting systematic reconnaissance and vulnerability scanning against organizations across Asia between late August and early October 2026. The report also describes LLMJacking — stealing LLM credentials and cloud resources from compromised enterprise environments and reselling access to high-performance AI models through underground proxy services.
The numbers behind the shift
Google’s data shows the pressure is already rising. Overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits — flaws vendors do not know about yet — rose from a 2025 average of 8 a month to 11 a month in 2026, then spiked to 22 in August alone.
The PaperCut case shows how fast the new cycle runs. In late August 2026, attackers exploited two PaperCut print-server flaws as zero-days — CVE-2026-81578, an authentication bypass, and CVE-2026-82078, an unsafe dynamic class loading bug — which chain together into unauthenticated remote code execution. Security researchers generated working exploits for the flaws within a day of disclosure, and GreyNoise reported a campaign four days after disclosure in which a threat actor likely used LLMs to build and test exploits, assemble a target list, and run a widespread campaign against hundreds of organizations.
What Google says comes next
Google places current activity at stage five of a seven-stage timeline it calls “minimal human-in-the-loop.” Its projections, as reported by BleepingComputer from the report’s timeline graphic, are that unsophisticated multi-agent attacks requiring little to no human interaction could arrive within about three months, followed by sophisticated, fully autonomous agentic attacks within about six months. Those timelines are forecasts, not observed activity.
There is one reassurance for defenders, for now. Google says these frameworks currently “trade stealth for velocity”: unrestricted autonomous execution generates noisy traffic, relies on recognizable identifiers, and leaves distinct code-level traces, so good logging and monitoring still catch them. The warning is that as the tools and the models behind them improve, both human involvement and detectability are expected to fall.
What this means for you
For anyone running a website or a business network, the practical change is the patch window. Treat critical, internet-facing updates — WordPress plugins and themes, print servers, VPNs, firewalls — as same-day work, not something for the next maintenance cycle. Turn on automatic updates where that is safe, and watch server logs for bursts of failed requests in the days after a patch release; that is what an automated diffing framework looks like from the defender’s side.
For everyone else, the advice is the same as ever, only more urgent: install operating system and app updates when they arrive, and do not leave old software running on an internet-facing device. The patch that fixes a flaw now also advertises it — and the advert is being read by machines.
Analysis: the patch window just became a race
Analysis: The old logic of patching was that a fix protected you and the attacker needed time to reverse-engineer it. Exploit_agent inverts that logic: the patch itself is the map, and the map is read at machine speed. The defence that still works is speed on the other side — patch fast, log everything, and assume any internet-facing system is being probed the moment its vendor ships a fix. Google’s “first observed” wording matters too: it means the first time Google saw it, not necessarily the first time it happened. Expect more frameworks like this to surface, and expect the three-to-six-month forecasts to be the numbers security teams plan around.
Frequently asked questions
What is exploit_agent?
Exploit_agent is a multi-agent AI framework that Google’s Threat Intelligence Group found running on an exposed staging server. It connects to several large language models, compares popular web software before and after security patches, drafts exploits for the flaws the patches fixed, and runs them automatically — with no human in the loop.
How does an AI agent turn a security patch into an exploit?
It uses differential analysis: it takes the code before a patch and the code after it, reasons about what changed and why, identifies the vulnerability the patch closed, then writes and fires an exploit aimed at systems still running the old version. The whole chain runs without a person approving any step.
Did exploit_agent hack the WordPress plugin in Google’s example?
Google does not say so. Its report shows the framework inspecting access-control code in the Spectra WordPress plugin, version 2.19.25, but it does not say a flaw was confirmed there or that any site running it was compromised. The plugin is an example of what the framework reads, not a confirmed victim.
How soon could fully autonomous cyberattacks become common?
Google’s timeline projects unsophisticated multi-agent attacks needing little to no human interaction within about three months, and sophisticated, fully autonomous attacks within about six months. Those are forecasts from the November 16, 2026 report, not attacks Google has already observed.
What should website owners do now?
Patch internet-facing software the day updates land, especially WordPress plugins, print servers, and VPNs. Enable automatic updates where safe, and monitor server logs for unusual bursts of requests after patch releases — automated frameworks generate noisy traffic that good logging still catches.
Sources:
- Google Cloud Blog — GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold (November 16, 2026)
- The Register — Google: AI agent turned a security patch into a working exploit with no human help (November 16, 2026)
- Cyber Magazine — Google Spots Fully Autonomous AI Attacks For the First Time (November 17, 2026)