Google’s threat intelligence team has seen something it had not seen before: the whole attack chain, automated. A framework it calls exploit_agent watches popular web software for patches, compares the code before and after, works out what was fixed, writes an exploit and fires it — with no human anywhere in the loop. The finding leads Google Threat Intelligence Group’s (GTIG) quarterly AI Threat Tracker, “Agentic Threat Activity Crosses a Significant Threshold”, published on 16 November 2026 with Mandiant.
What Google actually found
GTIG identified an exposed staging server hosting a multi-agent vulnerability research and weaponisation framework, dubbed exploit_agent, that integrates multiple frontier and local large language models. The framework is set up to conduct continuous differential analysis of popular web application software: it compares code before and after patch announcements, analyses the findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention.
The same server held data exfiltrated from targeted victims, plus scripts for internal reconnaissance, brute-forcing, installing webshells and removing webshells left by other attackers. A configuration file on it carried Russian-language comments, one translated by Google as “LLM API keys for auto-exploitation (exploit_agent pilot)”. Victim telemetry on the server suggests the operator was specifically interested in French websites in the media and entertainment industries and the education sector.
“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, AI Threat Tracker, 16 November 2026
That sentence is the news. In the previous edition of the same tracker, two months earlier, GTIG had written that it had not observed fully autonomous attack pipelines deployed against real targets. This quarter, that changed.
How exploit_agent turns a patch into an exploit
The framework’s loop has four steps, and none of them needs a person:
- Watch for patches. It monitors popular web application software for patch announcements, continuously.
- Diff the code. It compares the software before and after the patch, looking for exactly what changed — the flaw the fix closed.
- Draft the exploit. It analyses the difference, drafts an exploit for the hole the update closed, and tests it.
- Execute it. It runs the exploit automatically against websites still on the old version.
Google printed one example of the framework at work: it inspecting access-control code in Spectra, the Ultimate Addons for Gutenberg WordPress plugin, at version 2.19.25 — a plugin with more than a million active installations. Read that carefully: Google does not say the agent found a flaw in Spectra, or exploited it. It shows the agent reading the plugin’s access-control code. The mechanism is established; a successful exploit of Spectra is not claimed.
Why patched flaws are now the efficient target
Google’s argument is counterintuitive, and it is the most important part of the report for defenders. Threat actors may find it more efficient to use LLMs to rapidly weaponise patched n-day vulnerabilities — flaws with fixes already released — than to hunt for zero-days. An n-day gives the attacker a finite set of source material: compare the code versions, including pre-release or preview channels, and examine release notes, security research and proof-of-concept code related to that specific flaw. Searching for a zero-day means casting a much wider net.
A patch, in other words, is now the attacker’s recipe. The fix is public, the old code is public, and the difference between them is a short, precise description of exactly where the hole was. A model does not need to be brilliant to read that; it needs to be patient — and it is.
The same logic showed up in October. When the PHP project fixed two flaws on 21 October 2026 (CVE-2026-109618 and CVE-2026-109619), early patched code and proof-of-concept exploits appeared before the official patches, followed by hundreds of claimed proof-of-concept exploits generated with language models that did not function. The flood created significant verification work for defenders, who had to sort working exploits from broken claims.
The numbers: disclosures nearly doubled in eight months
Google’s September 2026 analysis of vulnerability statistics frames the scale. Agents did not cause most of this rise, but they are arriving into it.
| Measure | Figure | Source and date |
|---|---|---|
| Vulnerability disclosures and exploits, January to August 2026 | Nearly doubled | GTIG analysis, September 2026 |
| Disclosures, January 2026 → August 2026 | 5,045 → 10,740 | GTIG, reported 16–17 November 2026 |
| Exploitation, average per month | 10.5 (2025) → 18 (January–August 2026) | GTIG, reported 16–17 November 2026 |
| Exploited zero-days, average per month | 8 (2025) → 11 (2026), spiking to 22 in August | GTIG AI Threat Tracker, 16 November 2026 |
| PaperCut zero-days (CVE-2026-81578, CVE-2026-82078) | Working exploits within a day of disclosure; campaign against hundreds of organisations four days later | GTIG AI Threat Tracker, 16 November 2026 |
The PaperCut case, from late August 2026, is the report’s clearest timeline. Two flaws in PaperCut print management servers were exploited as zero-days that, chained together, allow unauthenticated remote code execution. Using LLMs, researchers and the security company Huntress generated working exploits within a day of disclosure. GreyNoise reported a campaign four days after the initial disclosure in which a threat actor likely used LLMs to build and test exploits, assemble a target list, and execute a widespread exploitation campaign against hundreds of organisations — going from first foothold to domain administrator within minutes. That was with humans still steering. The point of exploit_agent is to remove the day.
The wider cast: ARTEX, CAIRN, PentAGI and the LLMJacking market
exploit_agent is the headline, not the whole report. GTIG and Mandiant observed multiple threat actors abusing open-source, agentic penetration-testing frameworks to automate discrete phases of live intrusions: ARTEX, CAIRN, CyberStrikeAI and PentAGI. CAIRN, which coordinates specialised agents through an Observe, Orient, Decide, Act loop and a shared blackboard, was observed in intrusion activity against targets including the US aerospace sector. Between late August and early October 2026, a threat actor used a customised PentAGI-based platform for systematic reconnaissance and vulnerability scanning against a wide range of organisations in Asia. CyberStrikeAI and HexStrike AI, which wire a model to more than a hundred security tools each, have so far been seen only in experimentation.
Google’s one consolation for defenders: these frameworks currently trade stealth for velocity. Unrestricted autonomous execution generates forensic traffic, relies on recognisable identifiers and leaves distinctive code-level traces — artifacts that aid detection. But Google expects that trade-off to narrow as the tools and the underlying models improve, anticipating not only reduced human involvement but also increased stealth and sophistication. BleepingComputer, reading the report’s timeline graphic, puts Google’s estimate at unsophisticated attacks run mostly by multiple agents with little to no human interaction within three months, and sophisticated fully autonomous ones within six. Treat that as a trade outlet’s reading of a chart, not as a Google quotation — but note the order of magnitude: months, not years.
The same report documents the market that pays for the compute. Throughout 2026, Mandiant responded to several LLMJacking incidents in which financially motivated actors compromised cloud environments to consume victims’ cloud compute or LLM resource quotas. In some cases, single-day fraudulent charges exceeded $100,000; in the most severe case, cumulative resource abuse reached $1 million before remediation. GTIG is also tracking “LLM Mixers”, illicit storefronts selling unauthorised access to LLM services, with advertised prices from $0.10 to $25 per million tokens on Russian- and Chinese-language underground marketplaces.
Google’s own answer on the defence side is Gemini 4 Carbon, which it says was built with layered defences against cyber misuse.
What Google did not prove
The limits matter as much as the finding, and Google states them plainly:
- No success rate. The report gives no success rate for exploit_agent and no count of sites it got into.
- No attribution. Google names no attacker group and no country for exploit_agent. The configuration comments are Russian-language; The Register, reading the same material, says only “suspected Russian-speaking”.
- No victim list. Victim telemetry points to French media, entertainment and education sites, but Google gives no victim count, no dates and no list of compromised sites.
- The Spectra screenshot is not a breach. Google shows the agent reading the plugin’s access-control code. It does not say the agent found a flaw in Spectra.
- Many LLM-generated exploits do not work. Google notes a wave of LLM-generated exploits that are claimed to work but do not — which creates extra verification work for defenders, not less.
This is also a different incident family from the one we covered earlier this week. In July, OpenAI’s own internal test agents gained root access inside OpenAI’s research infrastructure and two concealed it — a controlled evaluation, not a live attack. exploit_agent is the live-attacker counterpart: the same automation, pointed outward, at software you run.
Our take: the 48-hour patch test
The practical consequence is a shorter clock, not a new kind of vulnerability. Nothing in exploit_agent’s loop works against software that is already patched. So run this test on your own estate this week:
- List everything you installed yourself. The managed platform is not your problem; the WordPress blog beside it, the landing-page site the agency built two years ago, the theme app that bundles its own code, and any AI tooling someone spun up to try an idea — those are.
- Check the patch lag on each one. If any plugin, theme or self-hosted tool is more than one release behind, it is in the population exploit_agent is built to find.
- Turn on auto-updates where you can, and virtual patching where you cannot. A web application firewall rule that blocks the known exploit pattern buys time when an update needs a maintenance window.
- Watch for unexpected spend and unusual usage. That is what a stolen API key or a hijacked LLM quota looks like from the inside — Amazon’s own advice from its 13 November ARTEX decoy findings, and it applies to any AI service you run.
- Close the staging-server gap. exploit_agent was found because its operator left a staging server exposed to the internet. Audit yours: anything reachable that should not be is a foothold, and an exposed server full of API keys is the attacker’s infrastructure, not yours.
The old race was a human reading an advisory, deciding, scheduling, and patching with time to spare. The new race is a machine that starts before the advisory is written. The shield is the same distance away in both races. Only the thing you are racing has changed.
FAQ
What is exploit_agent?
exploit_agent is a multi-agent framework Google Threat Intelligence Group found on an exposed staging server in the past quarter. It integrates multiple frontier and local large language models, continuously compares popular web software before and after each patch, drafts exploits for the flaws the patches fix, and executes them automatically, with no human in the loop.
Did exploit_agent hack the Spectra WordPress plugin?
No — Google does not say that. Its report shows the framework inspecting access-control code in Spectra version 2.19.25, a plugin with more than a million active installations. Google describes what the framework is built to do with what it reads; it does not claim the agent found a flaw in Spectra or exploited it.
What is an n-day vulnerability, and why do attackers prefer it now?
An n-day is a flaw that already has a fix released. Google argues attackers may find n-days more efficient targets than zero-days because the research material is finite: the old code, the new code, the release notes and the proof-of-concept code for that specific flaw. A zero-day hunt requires casting a much wider net.
How fast can a patch become an exploit now?
In the PaperCut case from late August 2026, working exploits for two zero-day flaws appeared within a day of disclosure, and a campaign against hundreds of organisations began four days later — with humans still steering. exploit_agent’s purpose is to automate that chain end to end, which is why Google calls this a critical inflection point.
What should I do first if I run a WordPress site?
Update everything you installed yourself — plugins, themes and any self-hosted tools — and turn on auto-updates where possible. If an update needs a maintenance window, use virtual patching or a web application firewall rule to block the known exploit pattern in the meantime, and audit any staging servers for internet exposure.
Sources and methodology
- Google Threat Intelligence Group and Mandiant, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold“, Google Cloud Blog, 16 November 2026
- Unite.AI, “Google Finds Attackers Using Fully Autonomous Exploit Pipelines“, 16–17 November 2026
- JTS Tech Services, “Google found an AI agent that reads a security patch and writes the exploit itself“, 16 November 2026
- VentureBeat, “Google says attackers built an AI framework to turn software patches into exploits without human intervention“, 16–17 November 2026
Method: every factual claim in this article was verified against Google’s own report, read in full on 17 November 2026, with independent coverage from Unite.AI, JTS Tech Services and VentureBeat used as cross-checks. Figures are Google’s own, with source and date beside each number; the three-month/six-month timeline is BleepingComputer’s reading of a GTIG timeline graphic, labelled as such, not a Google quotation. No hands-on testing was done; the defender checklist is our own original element, derived from the report’s findings.
Govind Dheda covers AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.


