Google says an AI agent turned patches into exploits, no human in the loop
Google Threat Intelligence Group found exploit_agent, an AI framework that turned security patches into working exploits with no human in the loop, a documented first.
Key points
- Google Threat Intelligence Group said it found a fully automated vulnerability exploitation system called exploit_agent, the first time the group has documented AI handling an entire attack chain without human involvement.
- The framework, linked to a suspected Russian-speaking operator, compared software code before and after security patches and then ran exploits automatically, with no human in the loop.
- GTIG said victim telemetry pointed to French media, entertainment, and education websites as the operator's targets of interest.
- Google researchers said threat actors are highly likely to keep using autonomous AI frameworks in attacks on organizations worldwide.
Google Threat Intelligence Group (GTIG) said it found threat actors using AI frameworks to fully automate vulnerability discovery and exploitation for the first time, with no human oversight at any stage. The activity centered on a framework called exploit_agent, found on an exposed staging server tied to a suspected Russian-speaking operator.
The Register, citing the Google findings, said the operator used the AI agent to convert security patches into working exploits before stealing data from victim organizations. Google placed the case in its fourth-quarter 2026 AI threat tracker and described it as a significant advance over what the group saw in the third quarter.
Key facts
- GTIG found exploit_agent on an exposed staging server. The framework connects to multiple frontier and local AI models and included Russian-language tooling.
- The system ran continuous differential analysis of popular web application software. It compared code before and after security updates to look for weaknesses.
- When it found a potential flaw, the framework drafted an exploit and ran it automatically. No human directed the process at any point.
- The same server held stolen data from victim organizations. It also contained scripts for internal reconnaissance, brute-forcing, installing webshells, and removing webshells left by other attackers.
- Victim data on the server pointed to French media, entertainment, and education websites as targets of interest.
- The report also covers other autonomous AI attack tools seen in active use: ARTEX, CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI.
Background
GTIG's prior quarterly reporting noted nascent adversary interest in agentic capabilities. The fourth-quarter findings mark a clear step forward: full automation with no human in the loop.
The appeal for attackers is efficiency. Searching for brand-new vulnerabilities requires a broad sweep across unfamiliar code. Focusing on patched vulnerabilities is more targeted. The research material is limited to two versions of a known codebase, plus release notes and security research. An automated framework can run that comparison continuously across many software packages at once. That shrinks the gap between a vendor releasing a patch and an attacker having a working exploit ready to deploy.
GTIG also documented several other autonomous multi-agent offensive frameworks active in the wild: CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI, alongside the previously reported ARTEX. The Register, citing GTIG, said PentAGI was deployed with proxy services in operations directed at the aerospace sector. Both financially motivated criminals and suspected espionage groups have been seen using these tools.
Google researchers noted that suspected espionage actors with a PRC nexus have been experimenting with these frameworks. The report covers several distinct threat groups, which shows the capability spreading across different types of actors and not confined to one geography.
What it means for you
Based on data found on the staging server, GTIG said the suspected operator focused attention on French websites in the media and entertainment industries and the education sector. The names of the affected organizations have not been made public. No official attribution beyond the description of a suspected Russian-speaking actor has been confirmed.
The aerospace sector also appeared in the findings. Any organization running internet-facing web applications faces a comparable risk. exploit_agent was built to analyze popular web application software broadly. That means the pool of potential targets is wide, not limited to any single industry or region.
The range of actors reinforces that point. The report covers financially motivated criminals alongside suspected espionage groups. These tools are not tied to one threat group, and the low cost of entry means more actors can adopt them going forward.
What you can do
The most direct protective step is to apply security updates promptly. exploit_agent worked by detecting the gap between a pre-patch and a post-patch version of web application software and then trying to exploit that opening. A system that is already updated gives the framework nothing new to act on.
For teams managing internet-facing applications, patching speed matters more than it once did. Automated frameworks can begin scanning against a newly released patch almost immediately. Waiting days or weeks to apply updates extends the window in which these tools can find and use a vulnerability.
Security teams should also check whether their monitoring can catch automated scanning activity. GTIG noted that fully autonomous frameworks tend to generate more network noise than human-directed attacks. Automated execution creates recognizable traffic patterns and leaves distinct code-level traces. That may change as the tools improve, but right now it is a real detection opportunity for defenders.
What happens next
Google researchers said threat actors are highly likely to keep using autonomous frameworks in campaigns against organizations worldwide. They cited the low cost of entry, the ease of adding new exploits, and the potential to scale broadly. As the underlying AI models advance, GTIG expects attacks to run with less human input and to produce traffic that is harder to separate from normal network activity.
FAQ
What did Google say was new?
GTIG said it was the first time the group had seen an AI agent handle both vulnerability discovery and exploitation from start to finish without human involvement. The Register, citing GTIG, described the agent as operating "with no human in the loop," calling it a first such occurrence.
What is exploit_agent?
It is a multi-agent AI framework found on an exposed staging server. The Register, citing Google, said the system compared software code before and after security patches, identified potential vulnerabilities, and then ran exploits automatically against the differences it found. The framework connects to multiple frontier and local AI models and included Russian-language tooling.
Who appeared to be targeted?
Victim data on the staging server pointed to French websites in media, entertainment, and education as the operator's areas of focus, GTIG said. The specific organizations involved have not been publicly identified. No official attribution beyond the description of a suspected Russian-speaking actor has been confirmed.
Sources
More from the network
Junious Media Network ↗Yellowjackets Season 4 Is the Show's Final Chapter
Showtime's survival drama Yellowjackets concludes after four seasons, with creators Ashley Lyle and Bart Nickerson saying the story is complete and no spinoff is planned.
junidaily.media ↗S&P 500 Drops Into Correction as AI Stocks Fuel Worst Week in 18 Months
AI-stock losses pushed the S&P 500 down 3.7% Friday, dragging the index into correction territory for the first time since March 2025.
junidaily.media ↗Macy's 100th parade will open with a 10-minute Broadway medley
NBC and Macy's will open the 100th Thanksgiving Day Parade on Nov. 26 at 8:30 a.m. with a 10-minute all-star Broadway medley introduced by Julianne Hough.
junidaily.media ↗Army Delays Iowa Lead Azide Blast After Toxic Plume Concerns
The U.S. Army postponed a controlled detonation of degraded lead azide in Iowa after residents and lawmakers raised safety concerns about a toxic lead plume.
junidaily.media ↗