P.K. SHARMA

Cyber security intelligence, AI governance, practitioner analysis

Google: one exposed server hosted a patch-to-exploit agent, no human in the loop; a relay says 'coming soon'

Google Threat Intelligence Group says it found one exposed staging server hosting exploit_agent, which compares code before and after patches, drafts exploits and runs them without a human. BleepingComputer's headline the same morning says such attacks are coming soon.

By · LinkedIn · · 18 min read

The report says 'for the first time'; the relay's headline says 'coming soon'

Google Threat Intelligence Group's report, dated 16 November 2026 on its page, says that 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". The evidence it gives is one exposed staging server hosting a framework Google calls exploit_agent. BleepingComputer's relay of the report, printed at 09:00 AM the same day, is headlined that fully automated AI attacks are coming soon. A finding written in the past tense and a headline written in the future describe one report. The first is about what GTIG found; the second is about what may follow.

The number that says where it bites is in the same report. Google's text describes exploit_agent as working from patches: it compares code before and after patch announcements. That is n-day work, on flaws that already have a fix. Zero-day exploitation, which has no fix to compare against, averaged 11 a month in 2026 against 8 in 2025, a rise of 37.5% (derived: 11 / 8 = 1.375), then reached 22 in August, 2.75 times the 2025 average and 2.0 times the 2026 average (derived: 22 / 8 and 22 / 11). The window an agent like this shrinks is the time between a vendor's patch, or a preview release, and a working exploit.

What that does not establish. The saved text does not say which flaws exploit_agent actually exploited, how many victims there were, how long the server was exposed, or when the first time was: GTIG's first identification is not a date of first use. It does not say whether the agent wrote its exploits unaided or an operator prepared inputs, which models it used, or who runs it, and this article names no actor. It does not say whether the 2026 average of 11 includes August's 22. And it does not contain the three month and six month projections that the relay attributes to Google.

Whose account this is. Google sells security (SecOps rulepacks, Mandiant consulting, Google AI Threat Defense and a programme for vetted defenders) and develops Gemini, whose misuse the report also describes. Every count, label and timeline is Google's own telemetry and Google's own wording; this site could test none of it. BleepingComputer is a news site relaying the report, and this article uses it only to show where its wording departs from the report's.

What Google says the staging server held

Google's text says GTIG "identified an exposed staging server hosting a multi-agent vulnerability research and weaponization framework, dubbed exploit_agent, that integrates with multiple frontier and local LLMs". The framework is set up to run continuous differential analysis of popular web application software. It compares code before and after patch announcements, analyses its findings for potential vulnerabilities, drafts exploits and automatically executes them without human intervention. Along with data exfiltrated from targeted victims, GTIG found multiple scripts for post-compromise activity: internal reconnaissance, brute forcing, installing webshells and removing other attackers' webshells.

Drawn from Google's report text. Step 1 comes from a general passage on n-days; steps 2 to 4 are one sentence about exploit_agent; steps 5 and 6 are on the same server.

Two phrases, one qualifier each. The finding is worded as "without a human in the loop", and execution as "without human intervention". Neither phrase says no human was involved in setting the framework up, choosing targets or reading the results. The report's own Table 1 shows a comment from an environment file; its Google translation reads as an instruction to a person to enter the API keys, and calls the set-up an exploit_agent pilot. The report also says the victim telemetry suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, and the education sector. That is a statement about an operator's interest. It is not a victim count, and the report gives none.

What Google's text prints about exploit_agent, claim by claim, and what each claim leaves open. Source: Google's report as read on 16 November 2026.

ClaimWhat Google printsWhat it leaves open
First timeIn the past quarter, GTIG identified full automation of discovery and exploitation without a human in the loop for the first timeWhen the framework first ran; how long the server was exposed
The serverAn exposed staging server hosting a multi-agent vulnerability research and weaponization frameworkHow GTIG found it; what happened to it
The modelsIntegrates with multiple frontier and local LLMsWhich models; whether Google's own were among them
The methodContinuous differential analysis of popular web application software; compares code before and after patch announcementsWhich software; which patches
ExecutionDrafts exploits and automatically executes them without human interventionAgainst whom; how often they worked
VictimsData exfiltrated from targeted victims; telemetry suggests interest in French media, entertainment and education sitesA count; which flaw; whether the agent obtained the data
ScriptsMultiple scripts capable of reconnaissance, brute forcing, installing webshells and removing other attackers' webshellsWhether the agent ran them; whether they were used
The operatorNothingAny actor, group or motive

Where automation bites: the patch, not the zero-day

Google gives its own reason why a patch-reading agent is the practical threat: threat actors "may find it more efficient to use LLMs to rapidly weaponize patched n-day vulnerabilities", because they can focus research on a finite set of source materials. The report says those materials include source code from pre-release or preview channels, release notes, security research and proof of concept code, and that "searching for zero-day vulnerabilities requires casting a wider net". It also says LLMs are shrinking the patch window for organisations; it prints no figure for that window.

Figures from one sentence of Google's report; the ratios are this site's arithmetic (derived). The report's Figure 3, the monthly counts as an image, was not read.

Reading the three figures. Google writes that, while overall disclosures and exploits nearly doubled between January and August 2026, zero-day exploits rose only marginally, from a 2025 average of 8 a month to 11 a month in 2026, before spiking to 22 in August. The derived rise from 8 to 11 is 37.5%, which this site does not call marginal or large: Google's word is its own. The two comparisons in that sentence are not the same kind. One runs from January to August 2026; the other compares a 2025 average with a 2026 average. The saved text gives no counts for the first, so no n-day share can be worked out from it. The text also does not say whether the 2026 average of 11 includes August's 22, or which months it covers.

One worked example of the window. Google says the two PHP flaws, CVE-2026-109618 and CVE-2026-109619, were discovered using LLM-enabled analysis, and that before the official patches of 21 October 2026 a proof of concept video and early patched code were published. That let researchers and attackers compare code versions before the patch was available. In the first days after the patch, hundreds of claimed but non-functional proofs of concept, likely generated with the help of LLMs, were published. Between that patch date and the report is 26 days (derived: 21 October to 16 November 2026). The text does not say that exploit_agent had anything to do with the PHP flaws. This site covered the PHP fixes in an earlier briefing.

Open source corroboration, as Google relays it. The report cites late August 2026 exploitation of two PaperCut print management flaws, CVE-2026-81578 and CVE-2026-82078, as zero-days that chained to unauthenticated remote code execution. It says security researchers and Huntress generated working exploits with LLMs within a day of disclosure, and that GreyNoise reported a campaign four days after disclosure in which a threat actor likely used LLMs, expanding to domain administrator within minutes. These are Google's summaries of other organisations' work. This site did not read the Huntress or GreyNoise pages, and the report does not tie the campaign to exploit_agent.

Where the relay and the report differ

BleepingComputer's piece is a secondary: it points to Google's report and uses it. These are the places where its wording departs from the saved report text. None is resolved here, and the relay is not wrong on every point: the report's text holds both the finding and a forward statement.

What each saved page prints. The red column is what neither establishes. The relay's estimates are quoted from the relay; they are not in Google's saved text.

The relay says Google's researchers estimate that within three months attackers can launch unsophisticated attacks needing "little to no human interaction" and that "sophisticated, fully autonomous" attacks are likely within six months. Neither figure is in Google's saved text. The report does have a figure titled LLM maturity timeline, and the relay shows a chart it credits to Google; both are images this site did not read, so this article says only that the saved text contains no three or six month figure, not that Google never wrote one.

Differences between the relay and Google's saved report text. Sources: BleepingComputer of 16 November 2026; Google's report as read on 16 November 2026.

What differsThe relay printsGoogle's saved text prints
Tense of the claimFully automated AI attacks are coming soon (headline)GTIG identified full automation in the past quarter, for the first time; and separately a forward statement that operations are approaching fully agentic intrusions
Dates for what followsThree months and six months, as Google's projectionsNo date in the saved text; Figure 2 is an image, not read
French websitesexploit_agent used in attacks targeting French media, entertainment and education sitesVictim telemetry suggests the operator was specifically interested in targeting them; data was exfiltrated from targeted victims; no count
What the server heldStolen victim data, reconnaissance information, webshells and brute forcing scriptsData exfiltrated from targeted victims, and scripts capable of reconnaissance, brute forcing, installing webshells and removing other attackers' webshells
Which frameworks run fully automatedAttackers systematically abuse ARTEX, CAIRN, CyberStrikeAI and HexStrike AI to run fully automated attacksFrameworks automate discrete phases and specific tasks in live intrusions; CAIRN and PentAGI seen deployed in the wild; CyberStrikeAI and HexStrike AI seen only in experimentation
The patch windowHas dramatically decreased the time defenders have to patchChaining these capabilities can dramatically reduce the window, and LLMs are shrinking it; no figure for the window
Claude CodeSuspected North Korean hackers using Claude Code to deliver malware; modified Claude CodeAfter intrusion, a suspected cluster modified .claude/settings.json files in internal repositories so that starting Claude Code triggered a backdoor download: the tool was the route, the backdoor was separate
The clusterA numbered cluster designation plus the name Midnight NeptuneOnly the name, as a suspected cluster; no numbered designation in the saved text
The PHP flawsNames the episode PHP2Shell; proofs of concept mostly AI generatedPrints two CVE ids, not that name; hundreds of claimed but non-functional proofs of concept, likely generated with the help of LLMs
How widespreadGTIG has observed its large-scale adoption by financially motivated and state-sponsored groupsThe summary says threat actors are increasingly leveraging LLMs for this; the evidence for full automation in the body is one staging server, plus GreyNoise's likely LLM use as Google relays it
The FBIMentions an FBI advisory last week on ARTEXNo mention of the FBI in the saved text. This site covered that advisory in its own briefing

What the report states, and what it does not

Questions a reader will ask, Google's answers and the gaps. Source: Google's report as read on 16 November 2026; the relay where named.

QuestionStatedNot stated
Did attackers fully automate discovery and exploitation?GTIG says it found a framework doing so, on an exposed staging server, for the first timeThe date of first use; any run against a named victim
Which flaws did it exploit successfully?It drafts and automatically executes exploits; Figure 4 shows it inspecting an access check in a WordPress plugin, version 2.19.25Any flaw, CVE or fix for that plugin; any success
How many victims?Data exfiltrated from targeted victims; telemetry suggests French media, entertainment and education sitesA count; whether the agent obtained the data
Which models?Multiple frontier and local LLMsNames; whether Google's Gemini was one
Who operates it?An operator whose interest the telemetry suggestsAny actor; no actor is named for exploit_agent
Was there no human at all?Google says execution is without human intervention; a comment in the environment file tells a person to enter the keysWho chose targets, who reviewed output
How much exploitation is LLM driven?Zero-days: 8 a month in 2025, 11 in 2026, 22 in August; overall disclosures and exploits nearly doubled January to August 2026Whether 11 includes August; the n-day count; any LLM share
When will fully autonomous attacks arrive?A forward statement without dates; the relay attributes three and six month estimatesWhere those estimates come from

What the report prints: names, one route, a comment and five CVE ids

Everything here is copied from Google's report as read on 16 November 2026. The report's text prints no indicators of compromise for exploit_agent or anyone else: no hashes, no IP addresses, no domains. Its figures are images, and this site did not read them, so it cannot say whether any figure prints one. This site did not visit, resolve or look up anything named in the report. The report also prints a block of PHP code from its Figure 4 showing an access check; this site does not copy it.

Text
exploit_agent
Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25
v21925.php
/wp/v2/spectra-popup
Names as printed in the report's text and the caption of its Figure 4, as read on 16 November 2026. The route string appears inside the code the report shows; the file name is in the report's comment above that code.

A reading, not a statement. The file name v21925.php and the version 2.19.25 carry the same digits once the dots are removed. That is this site's reading of the two strings. The report does not say the file is a copy of that version, does not call the access check exploitable beyond its own comment labelling it an insecure capability check, and names no CVE, fixed version or exploitation for the plugin.

The environment file comment. Table 1 of the report shows comments from an environment file that configures several LLM API keys for what Google calls autonomous exploitation. Below, the Russian lines and Google's English translation, as printed; the report's long dashes are shown as hyphens. A defender who finds a file like this on a seized or exposed host has found Google's description of the pilot's configuration, not an indicator of any victim.

Text
# LLM API-ключи для авто-эксплуатации (пилот exploit_agent)
# Впиши ключи сам - сюда, не в чат. Обе строки обязательны.
# LLM API keys for auto-exploitation (exploit_agent pilot)
# Enter the keys yourself - right here, not in the chat. Both lines are required.
Comments from the environment file in Table 1 of the report, with Google's translation, as read on 16 November 2026. Long dashes shown as hyphens.

The five CVE ids the report prints, as printed, with what Google says about each. This site did not check them against CVE.org or NVD (this site read only the saved page); the PHP pair also appears in this site's earlier briefing.

CVE id as printedProduct as Google describes itWhat Google says
CVE-2026-109618, CVE-2026-109619PHPDiscovered using LLM-enabled analysis; official patches 21 October 2026; early patched code and a proof of concept video public before the patch
CVE-2026-81578PaperCut print management serversAn authentication bypass flaw, exploited as a zero-day in late August 2026; chained with the next one for unauthenticated remote code execution
CVE-2026-82078PaperCut print management serversAn unsafe dynamic class loading vulnerability; same chain
CVE-2026-40217LiteLLMA sandbox escape that lets an attacker raise limited-access inference keys to full administrative credentials; exploited by a threat actor in an LLMjacking intrusion

The offensive frameworks the report names, with what Google says it has seen. Source: Google's report.

FrameworkWhat Google saysSeen how
ARTEXDescribed as a now highly publicised pentesting tool; rulepacks deployed in Google SecOps; the FBI's advisory on it is covered in an earlier briefingNamed among the frameworks threat actors use; the text does not say how GTIG observed it
CAIRNOpen source multi-agent framework with an observe, orient, decide, act loop and a shared blackboardObserved in the wild, including against targets in the aerospace sector in the US
PentAGIOpen source autonomous offensive framework; a customised platform based on it scanned organisations in Asia between late August and early October 2026Observed in the wild
CyberStrikeAI, HexStrike AIOpen source LLM-native and MCP-integrated offensive harnesses with over 100 and over 150 tools respectivelyExperimentation
Custom harnesses, ZeroHollow, Agent ZeroCustom harnesses; an inference service for exploit development and routing; an open source agent frameworkExperimentation observed

What else the report covers, in Google's labels

The report is wider than exploit_agent. These are Google's labels and counts, repeated once and attributed; none is verified here, and none is the point of this briefing. The agent configuration passage is the one a defender can act on at once: after intrusion, a suspected cluster Google calls MIDNIGHT NEPTUNE modified .claude/settings.json files in internal repositories so that starting Claude Code would trigger the download of a backdoor Google calls SOMBERMEME, and generated GitHub personal access tokens named after another vendor's assistant. The agent tool was the delivery mechanism; the malware was separate.

Other parts of Google's report, with Google's own labels. Source: Google's report as read on 16 November 2026.

TopicWhat Google saysGoogle's label or name
LLMjackingMandiant responded to several incidents in 2026; single-day fraudulent charges exceeded 100,000 USD; cumulative abuse reached 1 million USD in the most severe casesNo victim count; see this site's earlier briefing
Resale of stolen model accessPrices per million tokens from 0.10 to 25 USD, a 250-fold spread (derived: 25 / 0.10); storefronts offer monthly subscriptions or per-token pricesGoogle calls the storefronts LLM Mixers and names two vendors, APIMart and JetNight
State-linked use of LLMsAn Iranian government-backed actor uses LLMs including Gemini for reconnaissance and social engineering; PRC-nexus actors use them for reconnaissance and data analysisCALANQUE ION; a PRC contractor LEAKY CASTLE
Influence operationsGoogle assesses with high confidence that influence operations moved from experiment to full integration; one platform was engineered to manage over 100,000 accountsNo actor named in this briefing
Phishing as a serviceA Chinese-language service added an LLM chatbot to its June 2026 backend updateThe Jieyou phishing service
Google's own defencesRulepacks, a hardening guide, disabled assets, classifier updates, Google AI Threat Defense, a programme for vetted defendersGoogle AITD; the Fairwind Program

What to do, in order

Take this with you

This site's suggestions, in the order worth doing

  • Measure your own time from a vendor's patch to your deployed patch, in days, for every internet facing web application and plugin. An agent that compares a patch needs only the patch.
  • Start work when a patch is released. Do not wait for a CVE record or an exploit write-up: the report describes actors working from code and release notes.
  • Treat preview and pre-release channels as public source for this purpose. The report says actors compare code from them. Find out which of your suppliers publish them.
  • Verify a claimed proof of concept before you trust it or dismiss it. Google says hundreds of claimed but non-functional ones followed the 21 October patches, which is a verification burden on you. Test in an isolated lab and record the result.
  • Set billing and quota alerts on model API keys and cloud accounts, and rehearse revoking them. The report's single-day figure above 100,000 USD shows how fast a stolen key can run up a bill.
  • Review agent configuration files such as .claude/settings.json in repositories like any code that runs on start: require a reviewed change and alert on edits. This follows from the report's description of the DPRK-linked intrusion; it is this site's suggestion.
  • Ask your vendors how quickly they ship fixes, whether their pre-release code is public, and what they publish before a fix is available.

The question the report leaves open

Google's text puts one exposed server in evidence and says it was the first time. It does not say how many flaws that server exploited, how many victims it reached, or whether anything it drafted worked. What the text does say is how the agent works: it reads a patch and drafts an exploit, and its inputs, in Google's account, are code versions, release notes and public research. The relay's headline says the future is close; the report says GTIG found a first example on an exposed server. Either way, the number a defender controls is not in the report. How many days pass between a vendor's patch and yours, and who is measuring?

Key facts

Sources

  1. PrimaryGTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold, page data 16 November 2026: exploit_agent, the vulnerability statistics, the CVE ids, Table 1 and the mitigations. Figures 1 to 6 are images and were not read. The vendor's own account. Read by this site on 16 November 2026Google Cloud Blog (Google Threat Intelligence Group)accessed 2026-11-16
  2. Reported byNews relay printed 16 November 2026, 09:00 AM: Google warns that fully-automated AI attacks are coming soon. Attributes three and six month projections not in the saved report text. Used to show where its wording departs from the report. SecondaryBleepingComputeraccessed 2026-11-16

Share this briefing

Know someone who owns this problem? Send it to them.

Related briefings

The briefing, in your inbox

Practitioner analysis of cyber and AI security news. No vendor noise.

How often

Every new briefing in one email, at 7am, or at 7am, 12:30pm and 6pm. Nothing is sent when nothing is new. Unsubscribe any time.