🔍 Read the full analysis: The Evidence Behind Google’s Findings On AI Attacks Without Humans on ThorstenMeyerAI.com
Get hardware and tech essentials 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’s Threat Intelligence Group says it found an exposed server running exploit_agent, a framework that analyzed software patches, drafted exploits and executed them without a person approving each step. The finding concerns at least one observed system; Google’s broader forecast of attacks with little or no human involvement remains a projection, not an established norm.
Google’s Threat Intelligence Group says it found an exposed server running a framework that analyzed software updates, inferred the vulnerabilities they fixed, wrote exploits and ran them without human approval at each step. The finding, published with Mandiant on November 16, is evidence of at least one case of automated vulnerability exploitation; Google’s forecast that attacks with little or no human involvement will become more common is a separate projection.
In its report, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold,” Google describes a multi-agent system it dubbed exploit_agent. The framework connected to commercial and locally hosted large language models and, according to Google, compared software code before and after updates to identify possible security fixes, then drafted and executed attacks. Google called this the first time it had seen threat actors use language-model frameworks to fully automate vulnerability discovery and exploitation without a human in the loop.
The evidence described centers on one exposed staging server, not a documented count of successful intrusions. Google says the server also held data taken from victims and scripts for possible later activity, including searching inside networks and placing web backdoors. The victim data, Google said, suggested the operator was interested in French websites in media, entertainment and education. The report does not identify the operator or specify how many sites were compromised.
Google also describes open-source AI-assisted testing frameworks appearing in real intrusions. It says it observed CAIRN used against targets including the U.S. aerospace sector and a customized PentAGI platform used against organizations in Asia. These examples support Google’s account of attackers using AI tools, but they are distinct from the exploit_agent server and do not establish that every stage of those operations was autonomous.
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
What Automated Exploitation Changes
The report matters because exploit_agent, as Google describes it, links several steps that security teams have traditionally had to monitor separately: reviewing a software change, identifying a possible weakness, preparing an exploit and executing it. Automating that sequence could let an operator test or attack exposed systems faster than a person working manually. For defenders, the practical concern is response time: a vulnerable service may face automated probing soon after information about a patch becomes available.
That does not mean the report establishes a new, routine form of cyberattack. Google’s evidence is a specific server and framework, and the company does not report how often it succeeded. It also cautions that current tools may trade stealth for velocity: fast, noisy activity can leave detectable patterns. The finding is a reason for organizations to review patching, access controls and monitoring, not proof that every attacker can conduct an entire operation autonomously.
Products Worth Considering

Apple MacBook Air 13.6-inch Laptop with M5 Chip, 24GB Memory, 1TB SSD, Midnight

Trezor Safe 5 Crypto Hardware Wallet

LG 27G810A-B 27-inch UltraGear 4K UHD IPS Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Observed Tools, Projected Autonomy
Google’s report combines observed activity with a maturity scale for AI-enabled attacks. It labels the current stage as stage 5 of 7, describing activity in the wild with “minimal human-in-the-loop.” The next stages are projections: Google places attacks involving 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 time ranges are forecasts, not confirmation that such attacks have already become widespread.
This distinction helps explain differing headlines about the report. One account emphasized that automated exploitation had “arrived,” reflecting Google’s description of exploit_agent. Another emphasized that more fully autonomous attacks are “coming soon,” reflecting the future stages in Google’s timeline. The report supports the narrower claim that Google observed one framework automating discovery and exploitation; its broader prediction about the next stages remains prospective.
Google says its assessment draws on Mandiant incident-response engagements, global threat-actor tracking and live platform defenses. It also includes a figure showing the framework examining access checks in a WordPress plugin. The report does not say that this specific examination led to a successful attack, so it should not be treated as evidence of a confirmed compromise.
“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
Limits of the Server Evidence
Several key details remain undisclosed. Google does not name the server’s operator, state how many websites were successfully breached, or quantify the damage linked to exploit_agent. Its description of the French sites is based on victim data that “suggests” targeting interest; it does not establish that every site in those sectors was attacked. The report also does not confirm that the WordPress plugin analysis shown in its figure resulted in exploitation.
It is also unclear how reliably the framework operated across different software, whether a person set its targets or supervised activity outside the automated steps, and how often tools like this are being used. Google’s forecast of more autonomous attacks over the coming months is an assessment of expected development, not a confirmed schedule. The available evidence supports an observed instance, not a measure of prevalence across cybercrime.
Defensive Steps and New Evidence
Google advises organizations to prioritize actively exploited vulnerabilities, reduce the use of long-lived passwords and keys, restrict administrative access and watch for bursts of activity that are unlikely to come from a person working manually. These are defensive measures drawn from the report’s account of faster, potentially noisy tooling; Google does not claim they eliminate the risk.
The next significant evidence will be whether security researchers and incident responders identify additional cases, and whether those cases show attackers automating more of an operation than vulnerability discovery and exploitation. Google’s projected stages place greater autonomy within months, but the report does not provide a scheduled date for a further assessment. Until more cases are documented, the scope and reliability of these systems remain open questions.
Key Questions
What did Google say it found?
Google says it identified an exposed server running exploit_agent, a multi-agent framework that analyzed software changes, drafted exploits and executed them without a person approving each step, according to the report.
Does the report show that fully autonomous cyberattacks are widespread?
No. Google describes one server supporting automated vulnerability discovery and exploitation. Its broader prediction that attacks with little or no human interaction will emerge is a forecast, not evidence that such operations are already widespread.
Who was targeted in the activity Google described?
Google says data on the server suggested interest in French websites connected to media, entertainment and education. It did not identify the operator or give a confirmed number of compromised sites.
What can organizations do in response?
Google recommends prioritizing actively exploited flaws, limiting long-lived credentials, restricting administrative access and monitoring for unusually rapid bursts of activity. These measures can reduce exposure, but the report does not say they remove all risk.
Source: ThorstenMeyerAI.com
