🔍 Read the full analysis: What Google’s AI Attack Findings Show—and What Remains A Forecast on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Google says it found an exposed server running a framework that automatically analyzed software patches, developed exploits and executed them without a person approving each step. The company describes that as one observed case; its forecast that attacks will soon require little or no human interaction is a projection, not a report of attacks already operating that way routinely.
Google’s Threat Intelligence Group says it found an exposed server running software that could identify vulnerabilities from security updates, write exploits and execute them without human approval—a development the company calls its first observed case of fully automated vulnerability discovery and exploitation. The finding, published with Mandiant on November 16, 2026, is evidence that one part of an attack chain can already be automated; Google’s prediction that attacks with little or no human involvement will follow within months remains a forecast, not an observed general practice.
In its report, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold,” Google said researchers identified an exposed staging server hosting a multi-agent framework it dubbed exploit_agent. The system connected to multiple commercial and locally hosted large language models. According to Google, it monitored popular web software, compared code before and after updates, inferred which security weakness a patch addressed, drafted an exploit and automatically ran it.
Google said the server also held data taken from victims and scripts for possible follow-on activity, including searching within networks and installing web backdoors. Data on the server suggested the operator had targeted French websites in media, entertainment and education, the company said. Google did not identify the operator or report how many sites were compromised. Its report illustrates the framework examining a WordPress plugin’s access checks, but does not say that this particular examination resulted in an attack.
The report also describes AI-enabled tools used in real intrusions. Google said it observed CAIRN used against targets that included the U.S. aerospace sector, and a customized PentAGI platform used against organizations in Asia. These examples support the broader finding that attackers are using AI frameworks in operations. They do not establish that all stages of those attacks ran autonomously.
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
Automation Raises the Response Tempo
Automating the steps between a software patch and a working exploit could shorten the time defenders have to update vulnerable systems. A system that can inspect code changes and attempt exploitation at scale may help an attacker move quickly from a newly released fix to targeting systems that have not yet been patched. Google’s report makes that risk concrete by describing a framework that, in its account, carried out those steps without a person approving each one.
At the same time, Google said current tools “trade stealth for velocity”: they can act quickly but may produce activity that is easier to spot. That is a qualified assessment, not a guarantee that every automated attack will be obvious. For organizations, the report supports prioritizing flaws known to be actively exploited, limiting standing credentials and administrative access, and watching for unusual bursts of activity. These are defensive steps tied to the risks Google describes, not evidence that a particular organization has been targeted.
Products Worth Considering

Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation

Jackery HomePower 3000 Portable Power Station, 3072Wh, 3600W

LG 27GS60QC-B Ultragear 27-inch Curved Gaming Monitor (QHD, 180Hz, 1ms, 1000R)
As an affiliate, we earn on qualifying purchases.
One Case, Two Different Timelines
Google’s report separates observed activity from a seven-stage model of how attacker use of AI could develop. It places current activity at stage 5, “Minimal human-in-the-loop”, and marks later stages as projected. The company expects 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 time ranges are Google’s estimates, not confirmed dates for a change in attacker capabilities.
This distinction explains why coverage of the report can describe automated attacks as having arrived while also saying they are coming soon. The first phrasing refers to Google’s account of exploit_agent automating vulnerability discovery and exploitation in one observed case. The second refers to a wider pattern of attacks requiring little or no human interaction. Google’s current-stage label is not “no human”; its own framework reserves full autonomy for a projected later stage.
The tracker is Google’s quarterly account of how criminal groups, state-backed actors and influence operations use large language models. Google said its findings draw on Mandiant incident-response telemetry, global threat tracking and live platform defenses. The report’s evidence therefore combines a specific server discovery with observations of other AI-enabled tools in intrusions, while its stages beyond the present describe expected developments.
““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
Scope and Attribution Remain Unknown
The report does not say who operated the exposed server, how many websites were successfully compromised, or when the server was active. Google’s description of victim data as suggesting interest in French media, entertainment and education websites does not establish the full scope of targeting. Nor does the WordPress example confirm that the plugin analysis led to a successful intrusion.
It is also unclear how often exploit_agent or similar systems have completed an attack chain without people intervening. Google identifies one framework and other AI tools used in real intrusions, but does not provide a frequency estimate for fully automated exploitation. Its one-to-three-month and three-to-six-month timelines are projections; the report cannot confirm when, or whether, attacks with those levels of autonomy will become common.
Defenders Face an Uncertain Window
Google’s timeline points to possible growth in attacker autonomy over the coming months, but it is not a schedule of confirmed future incidents. The next useful evidence will be whether threat researchers document additional cases that automate more of the attack chain, and whether incident reports show these methods spreading beyond isolated systems or targets.
For now, organizations can act on the observed risk without treating the forecast as fact: prioritize patches for actively exploited flaws, reduce the lifespan and reach of credentials, restrict administrative access, and monitor for unusually rapid or high-volume activity. Google’s report does not identify a single expected next milestone or name a new finding date. Further changes in its assessment will depend on what researchers observe in subsequent investigations and incident-response work.
Key Questions
What did Google say it observed?
Google said it found an exposed server running exploit_agent, a framework that could analyze software updates, infer vulnerabilities, write exploits and execute them without human approval.
Does the report show that fully autonomous cyberattacks are widespread?
No. Google describes a specific observed case of automated vulnerability discovery and exploitation. It places current activity at “minimal human-in-the-loop” and treats broader, fully autonomous attacks as a future possibility.
What did Google say about the possible victims?
Google said data on the server suggested the operator was interested in French websites associated with media, entertainment and education. The report does not give a confirmed victim count or identify the server’s operator.
How soon does Google expect more autonomous attacks?
Google projects attacks with little to no human interaction in one to three months and fully autonomous attacks in three to six months. These are the company’s forecasts, not confirmed dates or evidence that such attacks will become common on that schedule.
What should organizations do now?
Google’s findings support prioritizing patches for actively exploited flaws, limiting long-lived credentials and administrative access, and monitoring for unusually fast or high-volume activity. The report does not establish that every AI-enabled attack will share the same detectable signs.
Source: ThorstenMeyerAI.com
