The patch is the exploit now: Google catches an AI agent weaponizing fixes with no human in the loop
Google threat researchers found an AI framework that reads security patches, works out the flaw each one fixes, and fires off exploits by itself. The patch window just collapsed.

On this page
In brief
- Google's threat researchers found a multi-agent framework, dubbed exploit_agent, that diffs code after patch announcements, drafts exploits, and runs them with no human in the loop.
- PaperCut, late August 2026: two chained flaws gave unauthenticated remote code execution; working LLM-built exploits appeared within a day, a campaign hit hundreds of organisations four days later.
- Zero-day exploits barely moved — 8 a month in 2025 to 11 in 2026 — while overall disclosures and exploits nearly doubled in January–August 2026. The growth is in n-days, not zero-days.
- The same report puts CAIRN in a US aerospace intrusion, LLMJacking bills over $100,000 in a day, and a North Korean cluster weaponizing Claude Code's settings file.
- The patch window is now measured in days, and the diffing never stops — internet-facing systems first.
For years, a security patch bought you time. The vendor found the flaw, shipped the fix, and the clock started: defenders raced to install it before attackers could work out what it fixed — and the patch itself was the safest document in the process, a description of the fix, not the flaw.
Google's Threat Intelligence Group says that clock has stopped working. In its fourth-quarter AI Threat Tracker, published Monday, GTIG reports the first case it has seen of threat actors fully automating vulnerability discovery and exploitation "without a human in the loop" — a framework that watches for patches, diffs the code, finds the flaw, writes the exploit, and fires it off. The patch is no longer the end of the story. It is the instruction manual.
The report, written by GTIG and Mandiant researchers, is the clearest picture yet of what "AI-powered attacks" actually looks like in practice. It is not a superintelligence finding zero-days. It is the boring, labor-intensive middle of vulnerability research — reading diffs, testing hypotheses, assembling exploit chains — handed to agents that do not sleep.
The first pipeline with no human in it
The centrepiece is a framework GTIG calls exploit_agent, found running on an exposed staging server: a multi-agent system that plugs into multiple frontier and local language models and runs continuous differential analysis of popular web application software. It compares code before and after patch announcements, identifies potential vulnerabilities from the differences, drafts exploits, and executes them automatically — no human intervention at any step.
The server told its own story. Configuration files on it carried Russian-language comments, including one labelling a set of LLM API keys as being for "auto-exploitation" in an exploit_agent pilot. Victim telemetry suggests the operator was specifically interested in French websites in the media and entertainment industries and the education sector. Alongside the framework, GTIG found data exfiltrated from targeted victims and scripts for post-compromise work: internal reconnaissance, brute-forcing, installing webshells — and removing other attackers' webshells.
GTIG is careful about what this is and is not. The previous quarterly report, from September, documented only partially automated scanning and exploitation. This is the step past that: the whole loop, closed. The group also notes the frameworks' weakness. Unrestricted autonomous execution is forensically noisy — recognisable traffic, identifiable markers, distinct code-level traces. The tools trade stealth for velocity. GTIG expects actors who need to stay quiet to use them selectively, and actors who care about scale — most cybercrime groups — to keep using them anyway, because the reach compensates for the detection risk.

PaperCut: the worked example
The best evidence for what this means in practice comes from late August 2026, when threat actors exploited two vulnerabilities in PaperCut print management servers as zero-days. Chained together, CVE-2026-81578, an authentication bypass, and CVE-2026-82078, an unsafe dynamic class loading flaw, allowed unauthenticated remote code execution. Researchers and the cybersecurity company Huntress generated working exploits for the flaws within a day of disclosure, using language models.
Four days after the initial disclosure, GreyNoise reported a campaign in which a threat actor likely used LLMs to build and test exploits, assemble a target list, and run a widespread exploitation campaign against hundreds of organisations. GreyNoise also highlighted the speed of execution: the actor expanded an initial foothold to domain administrator within minutes.
That sequence — disclosure, working exploit in a day, mass campaign in four, domain admin in minutes — is the shape of the new patch window. The human work that used to pace it, the reverse engineering and exploit development, is exactly the work these frameworks automate. The window is not only shrinking at the exploitation end, either. The PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619 were discovered using LLM-enabled analysis, and before the official patches were released on October 21, 2026, a proof-of-concept video and early patched code were already public — letting researchers and attackers alike compare code versions and attempt to build exploits before the patch was even available.
The numbers point at n-days, not zero-days
GTIG's September analysis, cited again in this report, explains why the focus has shifted. Overall vulnerability disclosures and exploits nearly doubled between January and August 2026. But zero-day exploits rose only marginally — from a 2025 average of 8 per month to 11 per month in 2026, before spiking to 22 in August. The growth is not in unknown flaws. It is in known ones, weaponised faster.
Hunting for a zero-day means casting a wide net across unfamiliar code. Weaponising an n-day — already patched and disclosed — means comparing two versions of the same code and reading the release notes. It is a finite, well-bounded job, and GTIG is blunt about the consequence: weaponising a patched n-day can be more efficient than hunting a zero-day, because the research starts from a known diff.
The rest of the report
The patch-to-exploit pipeline is the headline, but the tracker documents the same automation spreading across the whole attack lifecycle. GTIG observed the open-source multi-agent framework CAIRN used in intrusion activity against targets including in the US aerospace sector, and a customised platform built on the open-source PentAGI framework conducting systematic reconnaissance and scanning against a wide range of Asian organisations between late August and early October 2026.
The economics have industrialised too. Mandiant responded to several LLMJacking incidents through 2026, where actors compromised cloud environments to consume the victim's LLM quotas. Single-day fraudulent charges exceeded $100,000; in the most severe case, cumulative abuse reached $1 million before remediation. The stolen access is resold through "LLM Mixers," illicit storefronts that route buyer traffic through pools of stolen credentials and compute, with advertised prices from $0.10 to $25 per million tokens.
And the supply chain now runs through the AI tools themselves. In multiple intrusions against cryptocurrency organisations in North America and Europe, a suspected North Korean cluster tracked as MIDNIGHT NEPTUNE modified .claude/settings.json configuration files so that starting Claude Code would automatically download the SOMBERMEME backdoor — then masked its lateral movement by generating GitHub personal access tokens named after Microsoft Copilot. The coding assistant, configured to run code on the developer's behalf, became the delivery mechanism.
What changes for defenders
None of this changes what a patch is. It changes what the patch's release date means. When the interval between disclosure and weaponised exploitation was measured in weeks, patching promptly was good hygiene. When it is measured in days — and the diffing is done by a system that never stops watching for the next announcement — patching promptly is the whole defence, and the priority order changes with it: internet-facing systems first, because those are the ones the pipelines scan.
The honest caveat is GTIG's own: these frameworks are noisy, and noisy tools get caught. But operations will move further along the autonomous scale as the models improve — less human involvement, and eventually more stealth to go with the speed. The grace period defenders used to have was never guaranteed. It was a side effect of how slow the human work was. That side effect is now gone, and the patch notes are the first thing the other side reads.


