AI Attacks With No Human In The Loop: Google’s Evidence And Predictions
AIThis post was created with the assistance of artificial intelligence (AI).

🔍 Read the full analysis: AI Attacks With No Human In The Loop: Google’s Evidence And Predictions on ThorstenMeyerAI.com

Age 18–24?Offer from Amazon

Prime made for students and young adults

  • Fast, free delivery for dorm and study essentials
  • Prime Video and Amazon Music included
  • Member-only deals
Try Prime for Young Adults Free trial for eligible 18–24 year olds
As an affiliate, we earn on qualifying purchases.

TL;DR

Google’s Threat Intelligence Group says it found an exposed server running an AI framework that automatically analyzed software patches, developed exploits and executed them without a person approving each step. Google describes this as one observed case; its forecast that attacks with little or no human involvement will become more common in coming months is a separate, projected development.

Google Threat Intelligence Group says it found a server running an AI framework that analyzed software updates, inferred the vulnerabilities they fixed, wrote exploits and executed them without human approval at each step. The finding, published November 16 with Mandiant, is Google’s first reported case of attackers fully automating vulnerability discovery and exploitation; broader attacks with little or no human involvement remain a forecast, not an established trend.

Google said the exposed staging server hosted a multi-agent system it named exploit_agent, connected to several commercial and locally run AI models. According to the report, the framework compared software code before and after security patches, worked out what weakness a patch addressed, drafted an attack and automatically ran it. Google described this as the first time it had identified threat actors using large language model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop.

The server also contained data stolen from victims and scripts for possible follow-on activity, including searching inside networks and placing web backdoors, Google said. Information found there suggested the operator was interested in French websites in media, entertainment and education. Google did not identify who operated the server or say how many sites were compromised. Its report includes an example involving a WordPress plugin’s access checks, but does not establish that the examination resulted in a successful attack.

The report also describes other AI-assisted activity. Google said it had seen open-source penetration-testing frameworks used in real intrusions, including CAIRN against targets that included the U.S. aerospace sector and a customized PentAGI platform against organizations in Asia. These examples provide context for its assessment, but are distinct from the specific exploit_agent finding: the report does not say that every framework performed a fully autonomous attack.

At a glance
reportWhen: Report published November 16, 2026; for…
The developmentGoogle reported finding a multi-agent framework that automatically discovered and exploited software vulnerabilities, while predicting broader autonomous attacks within months.

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

  1. 1Prompting a chatbothuman-driven
  2. 2AI built into malwareautomated runtime call
  3. 3AI automation in operationsAI in the loop, humans in charge
  4. 4Experimenting with autonomous frameworksexperimental autonomy

Now · in the wild

  1. 5AI tools run several attack stages“minimal human-in-the-loop”

Projected · forecast, not observed

  1. 6Simple attacks with little to no human interactionnear-term, 1–3 months
  2. 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.

Forecast

Solid: observed by Google. Dashed: projected by Google.

Google’s own label for today: stage 5 of 7. Fully automated exploitation exists in at least one case. Fully autonomous attacks as the norm are still a forecast.

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

Automation Raises the Patch Stakes

Automating the path from a software patch to a working exploit could shorten the time between a vulnerability becoming visible and attackers trying to use it. A system that can repeatedly inspect updates and launch attacks may increase the speed and volume of activity defenders need to monitor. The reported server’s stored victim data and follow-on scripts also suggest that exploitation can sit within a wider intrusion process, although Google has not connected every item on the server to a confirmed compromise.

For organizations, the finding supports prioritizing flaws that are already being exploited, tightening administrative access and monitoring for unusually rapid bursts of activity. Google also says current tools can trade stealth for velocity: fast, noisy behavior may create opportunities for detection. That is a qualified assessment of current activity, not a guarantee that automated attacks will be easy to spot or stop.

Ninja CREAMi Scoop & Swirl 13-in-1 Ice Cream and Soft Serve Maker, NC701

Ninja CREAMi Scoop & Swirl 13-in-1 Ice Cream and Soft Serve Maker, NC701

Breville BES870BSXL Barista Express Espresso Machine, Black Sesame

Breville BES870BSXL Barista Express Espresso Machine, Black Sesame

As an affiliate, we earn on qualifying purchases.

One Case, Two Different Timelines

Google’s report, titled GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, is a quarterly assessment of AI use by cybercriminals, state-backed groups and influence operations. Google said it drew on Mandiant incident-response work, threat-actor tracking and its platform defenses. The central finding concerns one exposed server; the broader picture includes other AI tools and activity that Google says it has observed.

Google places the current state of activity at stage five of a seven-stage timeline, characterized as minimal human involvement. It projects attacks with little to no human interaction in the near term, defined as one to three months, and fully autonomous attacks in the medium term, defined as three to six months. Those stages are forecasts, not reports that such attacks are already common. The distinction explains why coverage of the report can describe automated exploitation as having appeared while also saying more autonomous attacks are expected soon.

The report’s wording matters: a framework automatically carrying out vulnerability discovery and exploitation does not mean that people played no role in selecting targets, setting it up or directing the broader operation. Google’s projected later stages describe a more extensive handoff, in which a person could set an objective and agents carry out more of the attack chain.

“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

Who Ran the Server Remains Unknown

Google has not named the operator of exploit_agent, established how many organizations were affected or disclosed whether the WordPress plugin analysis led to a successful intrusion. The report’s evidence comes from an exposed server; the material provided does not establish the full extent of its use or independently confirm every apparent victim connection.

It is also unclear how often attackers can deploy this kind of automated process successfully, whether it can operate quietly against well-defended targets, or how quickly the projected stages will develop. Google’s timeline is an assessment of expected capability, not a firm timetable for attacks to become routine. The single observed case should not be read as evidence that fully autonomous attacks are now standard practice.

Defenders Face a Near-Term Test

Google’s report sets out a forecast rather than a specific next milestone. It expects attacks with little to no human interaction in the one-to-three-month near term, followed by fully autonomous attacks in the three-to-six-month medium term. Whether those predictions materialize, and how broadly such systems are adopted, remains to be seen.

For now, organizations can act on the observed risk: prioritize actively exploited vulnerabilities, reduce long-lived credentials, restrict administrative access and watch for activity bursts that do not resemble normal human behavior. Those steps do not prove that an organization is facing AI-driven activity, but they address weaknesses that could make an intrusion easier as attacker tools change.

Key Questions

What did Google say it found?

Google said it found an exposed server running exploit_agent, a multi-agent framework that automatically analyzed patches, developed exploits and executed them. Google called it its first observed case of fully automated vulnerability discovery and exploitation without a human approving each step.

Does the report show fully autonomous attacks are common now?

No. Google describes one specific finding and labels the present stage minimal human involvement. Its expectation that fully autonomous attacks will emerge in the medium term is a forecast, not evidence that they are already common.

Who was behind the server, and how many victims were there?

Google has not identified the operator or given a number of affected websites. It said data on the server suggested interest in French media, entertainment and education sites, but did not establish the full scope of any compromises.

What should organizations do in response?

Google’s recommendations include prioritizing actively exploited flaws, limiting long-lived keys and passwords, restricting administrator access and monitoring for unusually rapid activity. These are defensive steps; the report does not say they can prevent every automated attack.

Source: ThorstenMeyerAI.com

You May Also Like

The Defender’s Counter-Cascade.

Google’s GTIG disclosed the first confirmed AI-built zero-day exploit, highlighting the deployment gap in AI-driven cybersecurity defenses as of May 2026.

Why Smaller AI Models Can Be Safer, Faster, and More Useful Than Giants

Outstanding small AI models offer safer, faster, and more practical solutions—discover why their advantages might surprise you.

The Overlooked Cost Of AI: Breaking Down The 176GB Memory Budget

Exploring the often overlooked memory components in AI inference, focusing on the 176GB weight size and the critical role of the KV cache and other factors.

A Skill Is A Folder, Not A Prompt: What Anthropic Learned Running Hundreds Of Them

Anthropic’s new approach packages AI knowledge into reusable folders called Skills, improving consistency and onboarding. This shift redefines AI agent design.