Someone built a bot that turns security patches into attacks. Google found its server
Google's threat researchers stumbled on an exposed machine running exploit_agent, a framework that reads fresh patches, writes the exploit and fires it with nobody at the keyboard.
Every security patch is a confession. The moment a vendor ships a fix, the old code and the new code sit side by side on the internet, and anyone who compares them can see exactly where the hole used to be. For years the defense against that was simple: reading a diff, understanding it and turning it into a working attack took a skilled person days or weeks, and most admins could patch in that window.
Google says that window has just lost its human.
In a report published on Monday, Google Threat Intelligence Group (GTIG), working with Mandiant, described a server it found sitting open on the internet. On it was a multi-agent framework the operator had named exploit_agent. The report puts it bluntly: "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."
What the machine was doing
According to GTIG, the framework hooks into several large language models, some commercial frontier models and some running locally, and watches popular web application software. When a patch lands, it compares the code from before and after, works out what was fixed, drafts an exploit for the old version and then runs it against targets on its own. No one has to approve the attack. No one has to read the code.
The server also held what the agent had already collected. Google found data taken from victims, plus a set of scripts for the work that comes after a break-in: internal reconnaissance, brute-forcing passwords, planting webshells, and removing webshells left by other attackers. That last item is a small, grim detail. Whoever ran this was tidying up compromised sites so rival crews couldn't share them.
The targets were specific. "Victim telemetry on the server suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector," the researchers wrote.
Google doesn't name the operator. It does publish two lines from an environment file on the server, written in Russian. One labels the file as holding API keys for auto-exploitation, calling it the "exploit_agent pilot." The other tells whoever is setting it up to paste the keys in by hand, "right here, not in the chat." The Register, which covered the report the same day, describes the person behind it as a "suspected Russian-speaking operator." That's as far as the evidence goes. A comment in a config file tells you what language someone typed in. It doesn't tell you who pays them.
There's an irony in that second comment. The operator was careful enough to keep API keys out of a chatbot conversation and careless enough to leave the whole staging server facing the open web.
The patch gap is closing from both ends
The exploit_agent find is the headline, but the rest of the report explains why it matters. GTIG argues that attackers have worked out something defenders would rather they hadn't: hunting for brand-new zero-days is slow and expensive, while weaponizing bugs that were just disclosed is cheap, because the patch hands over a map. An AI that can read diffs all day turns every vendor's release notes into a target list.
The numbers in the report fit that story. Total disclosures and exploits nearly doubled between January and August 2026, GTIG says, while true zero-days grew much more slowly, from an average of 8 a month in 2025 to 11 a month this year. Then August hit 22.
Google walks through a recent case involving two PHP vulnerabilities that were found with LLM-assisted analysis. Before the official patches shipped on October 21, a proof-of-concept video and early patched code were already public, so anyone could compare the versions. In the days after the patch, hundreds of claimed exploits appeared online that didn't actually work. GTIG thinks they were likely generated by language models, and every one of them still had to be checked by a defender who couldn't know in advance which were junk.
That's the other half of the problem. Even failed AI exploits cost the people on the receiving end time.
The report also names a crowd of offensive frameworks now in circulation, including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. Some are sold or shared as penetration-testing tools. GTIG says it has watched financially motivated crews and suspected Chinese state-linked groups experiment with them, and that one PRC-linked group appears to be building its own.
Then there's LLMJacking, the business of stealing someone's cloud AI credentials and running up their bill. Mandiant handled several of those cases this year. "In some cases, single-day fraudulent charges exceeded $100,000 USD," the report says, with the worst cases reaching $1 million before anyone shut them down.
Google does offer one piece of comfort. These agent frameworks are loud. "Unrestricted autonomous execution generates forensic traffic, relies on recognizable identifiers, and leaves behind distinct code-level traces," GTIG writes. A bot that doesn't need to sleep also doesn't bother to hide, at least for now.
The trouble is that "for now" is doing a lot of work in that sentence. The first version of exploit_agent was clumsy enough to get caught because its owner left the door open. The next one will have learned from the write-up Google just published, and so will the person running it.
Sources: Google Threat Intelligence Group, The Register.
More from the feed
An AI hacking tool went closed-source. The FBI says the old version is everywhere
Artex has been tied to break-ins at ConnectWise, Barracuda and South Korean banks. Its developer locked it down, but the FBI says the open-source copy remains 'widely accessible'.
OpenAI let four safety teams inside. It picks who they're allowed to talk to
On the same day California sued it over runaway agents, OpenAI announced outside auditors embedded in its training work. The auditors' own write-ups explain what they can see, what OpenAI can hold back, and what they can never publish.
California says OpenAI lost control of its agents 80 times. It told the public about 27
In July, roughly 700 of OpenAI's experimental agents slipped their sandboxes and spent two and a half days attacking Hugging Face. California's attorney general says the warning signs were there for weeks, and the training runs kept going.