$2 / month

Comment and reply across TML.

Start or manage membership

Most pulsed

Latest from other desks

tml.the modern leap
Tech / Cybersecurity

Attack Frameworks Now Run Intrusions Without Human Operators

Google's Threat Intelligence Group says it found a framework on an exposed server that reads patches, writes exploits, and fires them with no person in the loop. The claim is Google's, the telemetry is Google's, and the report also sells Google's defense. The mechanics still check out.

TML Tech DeskNov 17, 12:39 PM ET · 5 min read

Edited and fact-checked by Stephen Sprague

Missouri National Guard and Panamanian cyber responders work a joint cyber incident exercise in Panama City, Panama, on May 15, 2025 (archive). Photo: U.S. Navy photo by Petty Officer 1st Class Daniel James Lanari, public domain, via Wikimedia Commons.
Summary

GTIG says automated frameworks now turn patches into working exploits without human operators, which makes patch speed the defense that matters.

Tap a passage. React to the part that matters.

The staging server never clocked out. Around the clock, it compared freshly patched web software against the version from the day before the patch, worked out what the patch had fixed, wrote an exploit for the hole, and fired it. No operator sat in that loop.

Google's Threat Intelligence Group describes that server in its Q4 2026 AI Threat Tracker, published Monday, November 16. GTIG calls the framework exploit_agent. Analysts found it on an exposed staging server along with victim data and a toolkit for what comes after entry: internal reconnaissance scripts, brute-forcing tools, webshell installers, and a script for deleting other attackers' webshells. Configuration comments on the server were written in Russian. The victim telemetry GTIG recovered pointed at French websites in media, entertainment, and education.

Exploit_agent is one entry in a longer list. GTIG and Mandiant say they have now watched multiple actors run open-source agentic frameworks inside live intrusions. The report names ARTEX, CAIRN, CyberStrikeAI, PentAGI, and HexStrike AI. These are penetration-testing tools. They chain the phases of an intrusion (reconnaissance, exploitation, movement) into a workflow that large language models can drive with little human direction.

CAIRN runs its agents on an observe, orient, decide, act loop, with the agents sharing a common blackboard of findings. GTIG says it observed CAIRN used in intrusion activity against targets including the United States aerospace sector. A customized platform built on PentAGI spent late August through early October 2026 running systematic reconnaissance and vulnerability scanning against organizations across Asia. GTIG also reports a group it describes as PRC-nexus developing its own automated attack framework for likely espionage use.

The patch pipeline is the part with a clock on it. A security patch is a public document that names a weakness by fixing it. GTIG says actors now automate the comparison between patched and unpatched code, a technique called differential analysis, and use it to weaponize already-patched flaws (n-day vulnerabilities) at scale. The window between a patch shipping and an exploit existing has been shrinking for years. GTIG's claim is that the window now closes without a person deciding anything.

GTIG's own numbers give the claim a shape. The group counted vulnerability disclosures and exploits from January through August 2026 and found they nearly doubled against the comparison period. Zero-day exploits averaged 8 per month in 2025, then 11 per month in 2026, then 22 in August 2026 alone. Those figures are GTIG's, drawn from its tracking, and the report says so.

One August case shows the mechanics without any speculation. Attackers chained two flaws in PaperCut print management servers, CVE-2026-81578, an authentication bypass, and CVE-2026-82078, an unsafe dynamic class loading flaw, into unauthenticated remote code execution. Security researchers used language models to build working exploits within a day of disclosure. Four days after disclosure, a campaign GreyNoise tracked hit hundreds of organizations with exploits GTIG says were likely built the same way. Print servers are not glamorous targets. That is the point. The pipeline does not care what the software is. It cares that a patch described a hole.

GTIG is honest about the current limits, and the limits matter. Autonomous frameworks trade stealth for velocity. Unrestricted execution generates heavy forensic traffic, recognizable identifiers, and code-level traces that raise the chance of detection. A careful human operator is still quieter. The frameworks are faster, broader, and cheaper to run, and they do not need the operator to be skilled. GTIG's assessment is that adoption continues anyway, because scale is its own advantage.

The intrusion frameworks are not the only machinery in the report. GTIG says state-sponsored groups, commercial influence vendors, and political monetization networks now routinely use language models to produce synthetic media at scale, build localized personas, and automate distribution across platforms. Espionage groups use the same models for target research and data analysis. One supply-chain technique the report names deserves attention from any company whose engineers use AI coding assistants: attackers manipulate the configuration files of autonomous coding tools so the assistant itself drops a backdoor during ordinary local execution. The tool the developer trusts becomes the delivery mechanism.

The same report tracks a quieter business running on the same machinery. GTIG describes LLMJacking, in which financially motivated actors steal language-model credentials and cloud resources from compromised companies and resell access to high-performance models through underground proxy services the report calls LLM Mixers. The intrusion stack and the resale stack are the same stack. Access gets stolen by agents, then rented out by the hour.

Two cautions belong on the record. The first is sourcing. The claim that fully automated exploitation has happened without a human in the loop rests on GTIG and Mandiant telemetry. No independent firm has published its own confirmation of the exploit_agent server, and this story treats the finding as GTIG's, attributed, because that is what it is. The second is interest. The same report that announces the threshold also presents Google's remedy, an autonomous defense product called Google AI Threat Defense. A vendor can be right about a threat and still profit from the fear of it. Both things are in the document, so both things are in this story.

Strip the product out and a hard operational fact remains. Defense has always run on patch latency: how long a known hole stays open on your systems. That latency used to be measured against human attackers who had to read the patch, understand it, and build the exploit by hand. GTIG's report says the reader on the other side is now software, and software reads every patch the day it ships. An organization that patches monthly is no longer slow by human standards. It is slow by the only standard that now applies.

The conversation

A little more perspective.

Reading is free. Verified members can comment and reply for $2 a month.

The conversation.

Loading comments…

Keep your curiosity going.

Some links here are affiliate links. If you buy through one, we might make a commission — it never changes what you pay. As an Amazon Associate I earn from qualifying purchases.