A Guide To Google’s AI Attack Report: Findings, Forecasts, And Open Questions
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Google’s Threat Intelligence Group says it found a server running a framework that discovered and exploited software vulnerabilities without a person approving each step. Google places current activity at stage five of a seven-stage scale and forecasts attacks with little human involvement within months; those later stages have not been observed, according to the report.

Google’s Threat Intelligence Group (GTIG) says it found an exposed server running software that automatically identified and exploited vulnerabilities, marking the first case of this kind in its tracking. In a report published November 16 with Mandiant, Google said it expects attacks requiring little human involvement within months, while making clear that broader, fully autonomous attacks remain a forecast rather than an observed trend.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

The report, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold,” describes a framework Google calls exploit_agent. GTIG says the exposed staging server hosted a multi-agent system connected to commercial and locally run large language models. The framework monitored widely used web software, compared code before and after updates, inferred which flaws those updates addressed, then drafted and ran exploits without human approval at each step.

Google also found data it described as stolen from victims and scripts for later activity, including searching within networks and placing web backdoors. It said the data suggests the operator focused on French websites in media, entertainment and education. The report does not identify the server’s operator or state how many websites were compromised. It also shows the framework examining access checks in a WordPress plugin, but does not say whether that examination resulted in an attack.

GTIG says it has also seen open-source AI “pentesting” frameworks used in real intrusions. The report cites CAIRN activity against targets that included the US aerospace sector and a customized PentAGI platform used against organizations in Asia. Google’s account places current activity at stage five of seven, where AI tools perform several attack stages with minimal human involvement. The final two stages are projections.

At a glance
reportWhen: Published November 16, 2026; forecasts…
The developmentGoogle published a threat report describing one observed case of automated vulnerability discovery and exploitation and forecasting more autonomous attacks in the coming months.
Status board: what Google has seen, and what it only expects, from Google's threat report of November 16, 2026. Observed, highlighted as a first per Google: exploit_agent, one exposed server that found flaws, wrote exploits and ran them with no human in the loop. Also observed: AI attack frameworks in real intrusions, CAIRN against US aerospace targets and PentAGI in Asia; a PaperCut campaign with at least 440 installations at 395 organizations in 48 countries, per GreyNoise; stolen AI access with fraudulent bills above $100,000 in a single day; coding-assistant settings changed to download a backdoor. Forecast by Google: in 1–3 months, simple attacks with little to no human interaction; in 3–6 months, sophisticated, fully autonomous attacks. Not known: how many attacks already run fully on their own; what exploit_agent damaged and who runs it; from which date the forecast counts
What Google observed and what it only forecasts. Sources: Google Threat Intelligence Group, AI Threat Tracker, November 16, 2026; GTIG and Mandiant, October 30, 2026; GreyNoise, PaperCut campaign report. Graphic: Thorsten Meyer AI.

Automation Raises the Pace of Exploitation

Automating the path from a software update to a working exploit could shorten the time defenders have to assess and respond to newly disclosed flaws. If attackers can identify what a patch fixes and rapidly produce an attack, organizations may face pressure to prioritize updates and monitor exposed systems promptly. Google’s report, however, documents one server and one framework; it does not establish how common this capability is or how many successful intrusions resulted.

Google also says the current tools trade stealth for velocity: their activity is fast but noisy, which can make it easier to detect. That observation matters for security teams assessing the immediate risk. The report describes a developing capability, not evidence that all attackers can operate autonomously or that current defenses have become ineffective.

Products Worth Considering

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

LG UltraGear 27-inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDR True Black 400

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity

Bose TV Speaker (2nd Gen) Compact Soundbar with Bluetooth and HDMI Connectivity

As an affiliate, we earn on qualifying purchases.

Google’s Seven-Stage AI Scale

GTIG organizes attacker use of AI into seven stages. Its report says the first four describe earlier activity, from people prompting chatbots to AI assisting malware and operational tasks. Stage four covers experimentation with autonomous frameworks. Google labels stage five as current activity, with tools running several parts of an attack and people still involved to a limited degree.

Stages six and seven are forward-looking: stage six describes simple attacks with little or no human interaction, while stage seven describes sophisticated, fully autonomous attacks. Google estimates the first could emerge in the near term, one to three months, and the second in the medium term, three to six months. Those windows are forecasts in the report, not a record of attacks Google says it has already observed.

The report says its findings draw on telemetry from Mandiant incident response engagements, global tracking of threat actors and live platform defenses. Google presented the tracker as a quarterly review of how criminal groups, state-backed actors and influence operations use large language models. The stages are Google’s framework for describing that activity.

““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

Reach of the Exploit Agent Case

Google has not said who operated the exposed server, how many sites were compromised, or whether the WordPress plugin analysis shown in the report led to exploitation. Its description of French targets is based on data that it says suggests the operator’s interests; it does not provide a victim count in the supplied report material.

The broader trajectory is also unsettled. Google’s stage-six and stage-seven timelines are forecasts, not confirmed outcomes. The report does not establish how often fully automated exploitation is occurring across threat actors, how quickly the capability will spread, or whether future systems will retain the noisy behavior Google describes in current tools.

Watch for Broader Autonomous Activity

Google’s forecast points to the next one to six months as a period to watch for attacks involving less human interaction. The report does not give a specific date for a follow-up assessment or say what evidence would mark a shift from stage five to stages six or seven. Security teams will need further incident reporting to determine whether the activity remains isolated or becomes more widespread.

For now, Google’s recommendations, as summarized in the report, are to prioritize actively exploited vulnerabilities, reduce the use of long-lived credentials, restrict administrator access and monitor for bursts of activity that are unlikely to come from a person working manually. These steps address the observed speed and automation while the scale and direction of the threat remain unclear.

Key Questions

What did Google observe?

GTIG says it found an exposed server hosting exploit_agent, a framework that compared software before and after updates, inferred vulnerabilities and ran drafted exploits without a person approving each step.

Does Google say fully autonomous attacks are already widespread?

No. Google describes one case of automated vulnerability discovery and exploitation and places current activity at stage five on its seven-stage scale. It forecasts more autonomous stages but does not report them as established or widespread.

Who was targeted by the framework?

Google says data on the server suggests interest in French websites associated with media, entertainment and education. It has not named the operator or stated how many sites were compromised.

When does Google expect attacks with less human involvement?

Google forecasts attacks with little to no human interaction in one to three months and sophisticated, fully autonomous attacks in three to six months. These are projections from the November 16, 2026 report.

Source: Thorsten Meyer AI

You May Also Like

Quantum Sensors and Precision Navigation: Applications in Space and Defense

Beyond traditional methods, quantum sensors revolutionize space and defense navigation, unlocking possibilities that could change how we explore and protect our world.

The Lowdown On Running Frontier AI Models With Your Mac Studio

Explore how the new Mac Studio with 512GB memory enables local running of frontier-scale AI models, its capabilities, limitations, and implications for users.

Mesh LLM: Distributed AI Computing On Iroh

Mesh LLM introduces a distributed AI computing framework on Iroh, enhancing large language model deployment and scalability across decentralized networks.

Autonomous Robotics for Logistics and Home Assistance

Many autonomous robots are transforming logistics and home assistance, but how exactly do they adapt to complex environments and improve over time?