🔍 Read the full analysis: What Google’s AI Attack Report Says About Human Oversight on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Google Threat Intelligence Group says it found an exposed server running an AI framework that could identify vulnerabilities from software updates, write exploits and execute them without a person approving each step. Google describes this as a first observed case, but says attacks that operate with little or no human involvement more broadly are still projected, not established as routine.
Google’s Threat Intelligence Group says it found an exposed server running an AI-based framework that identified software vulnerabilities, drafted exploits and launched them without human approval at each step. The finding, published November 16 with Mandiant, is Google’s first observed case of attackers fully automating vulnerability discovery and exploitation; it does not establish that fully autonomous cyberattacks are now common.
In its report, titled GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, Google describes a multi-agent system it calls exploit_agent. The framework was hosted on an exposed staging server and connected to several commercial and locally run large language models. Google says it examined software code before and after updates, inferred what vulnerability a patch addressed, generated an exploit and automatically executed it.
Google also says the server held stolen victim data and scripts for follow-on activity, including inspecting networks and installing web backdoors. Data on the server suggested an operator was interested in French websites in media, entertainment and education. Google did not identify the operator or say how many sites were compromised. A report illustration shows the framework examining access checks in a WordPress plugin, but does not establish that this particular analysis led to an attack.
The report also discusses open-source AI-enabled penetration-testing frameworks that Google says have appeared in real intrusions. It cites CAIRN activity against targets including the U.S. aerospace sector and a customized PentAGI platform used against organizations in Asia. Those examples are separate from the exploit_agent server finding. Google says the tools currently tend to trade stealth for speed: their rapid activity can be conspicuous and easier for defenders to detect.
Google threat report · Nov 16, 2026
How far AI attacks have come, by Google’s count
Google sorts attacker use of AI into seven stages. Four are history, one is now, two are a forecast.
Google’s seven stages
Observed · end of 2025 – Sep 2026
- 1Prompting a chatbothuman-driven
- 2AI built into malwareautomated runtime call
- 3AI automation in operationsAI in the loop, humans in charge
- 4Experimenting with autonomous frameworksexperimental autonomy
Now · in the wild
- 5AI tools run several attack stages“minimal human-in-the-loop”
Projected · forecast, not observed
- 6Simple attacks with little to no human interactionnear-term, 1–3 months
- 7Sophisticated, fully autonomous attacksmedium-term, 3–6 months
Two headlines, one report
“Have arrived”
The Decoder, Nov 16
Follows one observed case: a framework that found and exploited flaws with no human in the loop.
Evidence · one server“Coming soon”
BleepingComputer, Nov 16
Follows the timeline: stages 6 and 7 are Google’s forecast.
ForecastSolid: observed by Google. Dashed: projected by Google.
Sources: Google Threat Intelligence Group, “GTIG AI Threat Tracker”, Nov 16, 2026 (Figure 2, LLM maturity timeline); GTIG and Mandiant, Oct 30, 2026; BleepingComputer and The Decoder, Nov 16, 2026. Stage names shortened by us.
Graphic: Powered by Thorsten Meyer AI
Why Automated Exploits Matter
The exploit_agent finding matters because it links AI-assisted vulnerability research to automated execution, rather than stopping at code analysis or suggested attack steps. If a framework can identify a newly patched weakness and act on it quickly, organizations may have less time to update exposed systems before attackers attempt exploitation. Google’s report does not quantify that time pressure or show how often this method has succeeded, so the operational scale remains uncertain.
For defenders, the report points to practical measures rather than a claim that human security teams are already obsolete. Google recommends prioritizing patches for flaws known to be actively exploited, reducing long-lived credentials, restricting administrative access and monitoring for unusually rapid bursts of activity. These steps address both the possibility of faster exploitation and the report’s observation that current automated tools may leave detectable patterns.
The distinction between a demonstrated capability and a projected future matters for risk decisions. A single observed framework shows that some attack steps can be automated; it does not show that attackers generally can conduct complex operations without people, evade detection reliably or achieve their objectives at scale.
Products Worth Considering

Logitech MX Master 4 Ergonomic Wireless Mouse with Haptic Feedback, Graphite

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)

Samsung Galaxy Tab A11+ Kids Edition, 128GB, Royal Blue
As an affiliate, we earn on qualifying purchases.
Observed Capability Versus Forecast
Google’s tracker is a quarterly account of how criminal groups, state-backed actors and influence operations use large language models. The company says its findings draw on Mandiant incident-response work, threat-actor tracking and live platform defenses. The report’s central finding is that, in the previous quarter, GTIG identified actors using LLM frameworks to automate vulnerability discovery and exploitation without a human in the loop.
Google places the current threat landscape at stage five of a seven-stage maturity scale, labelled “minimal human-in-the-loop.” It projects attacks with little to no human interaction in the near term, which it defines as one to three months, and fully autonomous attacks in the medium term, three to six months. Those intervals are forecasts in the report, not evidence that either stage has become widespread.
This distinction explains differing headlines about the report. Coverage describing automated attacks as having arrived refers to the exploit_agent case. Coverage saying they are coming soon refers to Google’s projections for broader attacks with little or no human involvement. Google’s own framework treats the present as limited human involvement, not a general state of fully autonomous operations.
“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.”
— Google Threat Intelligence Group
Scale And Operators Remain Unknown
Google has not disclosed who operated exploit_agent, how many victims were affected or whether the suspected targeting of French sites resulted in successful compromises. The report’s description of data on the server indicates what researchers inferred about the operator’s interests; it does not provide a confirmed victim count or identify the people behind the system.
It is also unclear how often the framework was used, whether it succeeded against the software it examined, and how much human involvement occurred outside the automated steps Google describes. The report identifies one server and a first observed instance, not a measurement of how prevalent this approach is across attackers.
Finally, Google’s one-to-three-month and three-to-six-month timelines are projections. The report does not establish that attacks with little or no human involvement will arrive on those schedules, or that such attacks will be effective, stealthy or routine once developed.
Defenders Face A Shorter Patch Window
Google’s immediate guidance is for organizations to patch actively exploited vulnerabilities first, shorten the lifespan of credentials, limit administrative access and look for unusually rapid activity that may indicate automation. These are defensive recommendations from the report, not evidence that a specific new attack is underway against every organization.
The next measure of progress will be whether researchers observe more real-world intrusions in which AI systems perform multiple stages with little human direction. Google’s report gives estimated windows for that development but does not announce a fixed follow-up date or a confirmed next milestone. Until additional cases are documented, the clearest distinction remains between one observed automated exploitation framework and broader autonomy that Google expects could develop.
Key Questions
What did Google say it observed?
Google says it found an exposed server running exploit_agent, a multi-agent framework that analyzed software changes, inferred vulnerabilities, generated exploits and executed them without a person approving each step.
Does the report say fully autonomous cyberattacks are common now?
No. Google reports one observed case of automated vulnerability discovery and exploitation. It labels the current stage “minimal human-in-the-loop” and presents broader, fully autonomous attacks as a forecast.
Who was behind exploit_agent, and how many sites were affected?
Google has not identified the operator or provided a victim count. It says data on the server suggested interest in French media, entertainment and education websites, but does not confirm how many were compromised.
What does Google recommend organizations do?
Google advises prioritizing patches for actively exploited flaws, reducing long-lived credentials, restricting administrative access and monitoring for unusually rapid activity that could indicate automated operations.
Source: ThorstenMeyerAI.com
