Google announced “The era of autonomous attacks has arrived” in a headline on Monday. Its investigators documented a narrower breakthrough. On an exposed server they found software that could identify weaknesses and attempt to exploit them without someone directing each step. A person still set the operation up. Google’s report left both the operator’s identity and the software’s success rate unknown.
The distinction matters. An investigation into what ran on one server can show that a stage of an intrusion has changed. It leaves open how often the software succeeds and how much of the rest of an attack still needs a person. Google used the broader headline in a separate company post. By Monday evening, a Polymarket post was circulating it inside quotation marks.
Google declares “The era of autonomous attacks has arrived,”
What Google documented
Google’s Threat Intelligence Group and Mandiant say they found an unattended system for finding and exploiting security flaws, which their November 16 report calls exploit_agent. The investigators found victim data on the server. They did not report how many sites the tool compromised or how often it succeeded. In September, the same research group had already reported intrusions that automated several stages with little human involvement.
“For the most part, attackers aren’t using new groundbreaking methods,” John Hultquist, the group’s chief analyst, said in Google’s companion post. The report says automated activity can be more conspicuous and unreliable. It also lets attackers do more work in less time. On November 13, the FBI said some intrusions involving a separate framework called ARTEX reached elevated access within three hours.

The customer has a clock
For a company whose public systems have an unfixed flaw, a tool can be clumsy and still matter. What counts is whether the company can apply a vendor’s fix before someone reaches the exposed system. Verizon’s 2026 Data Breach Investigations Report found that vulnerability exploitation was the initial means of access in 26 percent of the small and midsize business breaches it studied. Defenders already faced that pressure when Cisco disclosed an actively exploited firewall flaw in October.
There is an obvious market for software that tells companies which flaws to fix first and how quickly they can fix them. Google introduced AI Threat Defense in May and promoted it again alongside Monday’s warning. It combines technology from Gemini, Wiz and CodeMender with Mandiant’s threat intelligence. Alphabet’s filings put the recorded purchase prices of two businesses behind that effort, Wiz and Mandiant, at about $35.6 billion combined. Google can document a real change in attacks and have a large financial interest in what companies buy to address it.

A small employer can ask what the software would buy in response time. A tool that flags a vulnerability will be worth more if the company already knows who can take a service offline, approve an update and verify that it worked. The same software may deliver a longer list of warnings to an organization that has nobody authorized to make those decisions. Boards can start by asking how much time passes between a critical update and its installation.
What will change first
Google’s report projects simpler attacks with little or no human involvement in one to three months and sophisticated, fully autonomous attacks in three to six months. Our prediction is that within six months a named victim will disclose an incident with enough evidence to show that software ran the central stages on its own. A disclosure with a timeline and an account of which decisions people made would give readers something that the November report and its headline leave unsettled.
6% chance the U.S. enacts an AI safety bill by end of year.
Federal law is likely to move more slowly. Congress has frontier-model oversight proposals in committee. The June executive order directs a voluntary framework for developers to give the government early access to certain models. On Monday evening, Polymarket priced the chance of a federal law meeting its AI safety criteria by year-end at 6 percent. A recorded committee vote or a bill that gains a place on the legislative calendar would be more informative than another few points of movement in that market.
Common questions
What did Google report on November 16?
Google reported that it had observed software finding and attempting to exploit security flaws without direct human input.
What is exploit_agent in Google’s report?
Exploit_agent is Google’s name for software it observed automating a stage of a cyberattack.
Who operated the software Google found?
Google did not identify the operator of the software it described on November 16, 2026.
How many victims did Google confirm?
Google reported finding victim data but gave no count of sites compromised by that software.
Were the reported attacks fully autonomous?
No. Google documented one stage running unattended. Someone still configured the tool.
How quickly did the FBI say ARTEX intrusions moved?
The FBI said on November 13 that some ARTEX intrusions reached elevated access in under three hours.
How often did attackers exploit a vulnerability to get in?
Verizon’s 2026 report found vulnerability exploitation was the first form of access in 26 percent of sampled small and midsize breaches.
When did Google introduce AI Threat Defense?
Google introduced AI Threat Defense in May 2026 and promoted it alongside its November 16 warning.
What does Google AI Threat Defense combine?
Google AI Threat Defense combines Gemini, Wiz and CodeMender technology with Mandiant’s threat intelligence.
How much did Alphabet record paying for Wiz and Mandiant?
Alphabet’s filings give combined recorded purchase prices of about $35.6 billion for Wiz and Mandiant.
What measure can boards ask about first?
Boards can ask how many hours pass between the release and installation of a critical security update.
What timeline did Google give for more advanced attacks?
Google projects sophisticated, fully autonomous attacks in three to six months in its November 2026 report.
What odds did Polymarket price for federal legislation?
On November 16, Polymarket displayed 6 percent for a law meeting its AI safety criteria by year-end.



