TL;DR
Google Threat Intelligence Group says it found an exposed server running software that automatically analyzed security updates, developed exploits and executed them without a person approving each step. The report describes this as one observed case; broader attacks with little or no human involvement remain a forecast, and Google says current tools can be fast but noisy.
In its AI Threat Tracker, published with Mandiant, Google said the server hosted a multi-agent framework it named exploit_agent. The system compared software code before and after updates, inferred which security weaknesses had been fixed, drafted exploits and automatically ran them, according to the report. It could connect to multiple commercial and locally run large language models.
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
Google said the server also held stolen victim data and scripts for later activity, including exploring networks and installing web backdoors. Data on the server suggested interest in French websites in media, entertainment and education, but Google did not identify the operator or report how many sites were compromised. A sample in the report showed the framework examining access controls in a WordPress plugin; Google did not say whether that examination resulted in an attack.
The report also describes attackers using open-source AI penetration-testing frameworks in real intrusions. Google cited CAIRN activity targeting, among others, the U.S. aerospace sector, and a customized PentAGI platform used against organizations in Asia. These examples show AI tools being used in attack operations, but they are distinct from the exploit_agent case involving automated flaw discovery and exploitation.

A Faster Path From Patch to Exploit
Software updates can reveal what a fix changed, giving attackers clues about the weakness it addressed. A system that rapidly compares code, develops an exploit and launches it could shorten the time defenders have to respond. Google’s finding gives security teams a concrete example of that capability, though the report does not establish how widely it has been used or how many victims it reached.
Google also says current AI-enabled tools trade stealth for velocity: their rapid activity can be noisy and easier to detect. That is a qualified assessment, not a guarantee that future tools will remain conspicuous. The report recommends prioritizing actively exploited flaws, reducing long-lived credentials, restricting administrator access and monitoring for unusual bursts of activity.
Products Worth Considering

Jackery Solar Generator 1000 v2 with 200W Solar Panel, 1024Wh LiFePO4 Battery

Bose QuietComfort Wireless Noise-Cancelling Headphones, Sandstone

Jackery Explorer 300 Portable Power Station, 292Wh
As an affiliate, we earn on qualifying purchases.
Google’s Seven-Stage Threat Timeline
Google places current activity at stage five of seven in its timeline: tools run several parts of an attack with minimal human involvement. Its earlier stages cover human-directed prompting, AI features built into malware, AI-assisted operations and experiments with autonomous frameworks. Google says those earlier uses had been observed by the report’s cutoff.
The final two stages are projections, not reported observations. Google estimates that simple attacks with little or no human interaction could appear in the near term, one to three months, and that sophisticated fully autonomous attacks could follow in the medium term, three to six months. The estimates describe Google’s expectations, not confirmed milestones.
““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
The Scale of Exploit Agent Activity
Google did not say who operated the server, how many organizations were affected or whether every exploit the framework drafted was successful. Its report links data on the server to likely interest in French websites, but does not establish the full extent of targeting or confirm that the WordPress plugin check led to compromise.
It is also unclear how often attackers have used comparable systems. Google’s stage-six and stage-seven timeline is a forecast; the report does not establish that simple, low-interaction attacks or sophisticated fully autonomous attacks have become routine. The single server finding supports a claim about one observed capability, not its prevalence.
Defenders Watch for Faster Attacks
Google’s report points organizations toward patching vulnerabilities already under active exploitation, limiting long-lived passwords and keys, tightening administrator access and looking for bursts of activity inconsistent with manual work. These measures address risks the report describes without assuming every attack is autonomous.
Google’s forecast windows place the next possible developments over the coming months. Whether those estimates are borne out, and whether AI-enabled attacks become quieter or more widespread, remains to be seen. The report does not name a specific next milestone or provide a date for a follow-up assessment.
Key Questions
What did Google observe?
Google says it found an exposed server running exploit_agent, a framework that analyzed software updates, developed exploits and automatically executed them without human approval.
Does the report say fully autonomous attacks are common?
No. Google reports one server case involving automated vulnerability discovery and exploitation. It places current activity at stage five of its seven-stage timeline and presents broader, more autonomous attacks as forecasts.
Who was targeted by the server operator?
Google said data on the server suggested interest in French websites in media, entertainment and education. It did not identify the operator or say how many sites were compromised.
When does Google expect attacks with less human involvement?
Google forecasts simple attacks with little or no human interaction in one to three months and sophisticated fully autonomous attacks in three to six months. Those timeframes are predictions, not confirmed events.
What steps does Google recommend for defenders?
The report recommends prioritizing actively exploited flaws, shortening the life of passwords and keys, restricting administrator access and monitoring for unusually rapid bursts of activity.
Source: Thorsten Meyer AI
