Google Threat Intel: AI Agents Now Turn Security Patches Into Exploits
Google Threat Intelligence Group says attackers automated vulnerability discovery and exploitation with no human in the loop, and found a framework called exploit_agent running on an open staging server.

Why it matters
- GTIG says it identified threat actors fully automating vulnerability discovery and exploitation "without a human in the loop" for the first time, in its AI Threat Tracker released Nov. 16, 2026.
- August 2026 saw 22 zero-day exploits, against a monthly average of 11 from January to August 2026 and 8 in 2025.
- Working PaperCut exploits for CVE-2026-81578 and CVE-2026-82078 were built with LLMs within a day of disclosure; GreyNoise saw a campaign hit hundreds of organizations four days after disclosure.
- PHP's Oct. 21 fixes for CVE-2026-109618 and CVE-2026-109619 were followed within days by hundreds of claimed, non-working proof-of-concept exploits.
- GTIG projects sophisticated, fully autonomous agentic attacks three to six months out, and prices stolen model access at $0.10 to $25 per million tokens.
Google Threat Intelligence Group says it has seen threat actors use language-model frameworks to fully automate vulnerability discovery and exploitation "without a human in the loop" for the first time. Its quarterly AI Threat Tracker, released Monday, Nov. 16, documents the toolkit behind that claim.
The case starts with an accident. A staging server left open to the internet gave Google's researchers a direct look at the thing security teams have long feared: software that reads a vendor's security patch, works out which flaw it closes, writes an exploit and runs it with nobody steering.
What is exploit_agent, and how does it work?
Google names the framework exploit_agent. It plugs into several frontier and locally run language models, then runs a loop. It compares popular web application code before and after patch announcements, flags likely vulnerabilities, drafts exploits and executes them.
The loop is continuous, so a patch release doesn't end the hunt. It may start it. Every announcement of a fixed flaw hands the system two versions of the same code to compare, which is why GTIG argues the diff is such a productive input for models.
The server held more than the agent. Google found stolen victim data on it, next to scripts for internal reconnaissance, brute-forcing and installing webshells. One script removed webshells left behind by other attackers, which suggests the operator shared the box with human intrusions running at the same time.
Victim telemetry pointed to French websites in media, entertainment and education.
The configuration file's comments are in Russian. In Google's translation, one labels the API keys "for auto-exploitation (exploit_agent pilot)" and another tells whoever sets it up: "Enter the keys yourself — right here, not in the chat."
That last instruction is the telling detail. Someone provisioned the keys by hand; the exploitation loop after that ran on its own. The operator still exists, but the operator's job shrank to pasting credentials into a form.
Why do patches make better targets than zero-days?
For IT teams, the problem is timing. Diffing a patch to find the bug underneath has long been a specialist's job, and specialists don't scale across thousands of vendor advisories a week.
GTIG says attackers may find it more efficient to point language models at flaws that were just fixed than to hunt for zero-days, because the material is finite. Two versions of the code, the release notes and any published proof-of-concept code.
Zero-day hunting asks an attacker to find the flaw before the vendor does. Patch analysis asks the attacker to look at a flaw the vendor already documented, which is a much shorter search.
How fast did the PaperCut exploits show up?
PaperCut is the case Google cites. Attackers exploited two flaws in its print management servers, CVE-2026-81578 and CVE-2026-82078, as zero-days in late August.
Researchers and the security firm Huntress used LLMs to build working exploits within a day of disclosure. GreyNoise later reported a campaign, four days after disclosure, in which an attacker likely used LLMs to build exploits and hit hundreds of organizations. That attacker went from an initial foothold to domain administrator within minutes, GreyNoise found.
The broader counts move in the same direction:
- Vulnerability disclosures and exploits nearly doubled between January and August, by GTIG's count.
- Zero-day exploits per month averaged 8 in 2025 and 11 from January to August 2026.
- August 2026 alone recorded 22 zero-day exploits, per the AI Threat Tracker chart.
- Zero-days rose more slowly through the period, then jumped in August.
Four days from advisory to hundreds of victims is the number security managers should hold onto. It compresses the window in which a patched system stays unpatched in practice.
Do the AI-written proof-of-concepts actually work?
Not always, and that is still a problem. PHP shipped fixes for CVE-2026-109618 and CVE-2026-109619 on Oct. 21. Within days, hundreds of claimed proof-of-concept exploits appeared, which Google says didn't work and were likely written with LLM help.
Defenders still had to verify or debunk each one. In other words, the work moved from the attacker to the person reading the queue. A flood of plausible-looking but broken exploit code consumes the same triage hours as a flood of working exploit code, and only one of the two arrives with a body count.
What are the multi-agent frameworks doing in the wild?
The exploit_agent find is the sharpest example in a report that tracks attackers moving from typing prompts to running open-source, multi-agent pentesting frameworks.
Besides ARTEX, the tool the FBI flagged in a FLASH alert last week, GTIG has seen CAIRN and PentAGI used in the wild. CAIRN turned up in intrusions against targets that included the U.S. aerospace sector.
Those frameworks "currently trade stealth for velocity," GTIG wrote. They leave forensic traffic, recognizable identifiers and code-level traces that raise the odds of detection. The trade is deliberate. A framework that runs fast and leaves fingerprints is worth more to an attacker than a quiet one that takes ten times longer to reach the same host.
That trade is the defensive opening. Google Security Operations has added rulepacks that detect ARTEX, and Mandiant has published a hardening guide built around the speed of LLM-driven attacks. Both bets assume the noise continues.
The report also updates Google's tracking of LLMJacking, the theft of cloud AI credentials that surfaced alongside the Fireworks AI breach earlier this month. Underground resellers price stolen model access at $0.10 to $25 per million tokens, GTIG found. That range sets the input cost for every automated loop described in the report.
What's the timeline for autonomous attacks?
A timeline in the report projects that multi-agent attacks with "little to no human interaction," though still unsophisticated, are one to three months away. "Sophisticated, fully-autonomous/agentic attacks" are three to six months out.
The two-stage shape matters. Unsophisticated does not mean harmless; it means the tooling is fast and the payloads are sloppy, which favors defenders who can read logs faster than the attacker can run tooling.
So what: the next quarter tests whether vendor triage capacity can keep up with machine-generated exploit code, and GTIG's own one-to-three-month projection leaves security teams no slack to spare.
Source: webpronews.com
More from Sophie Lindqvist
Show full bio
Staff writer covering regulation, policy and their impact on artificial intelligence and machine learning at AI In Context.
749 articles
Related articles
- Google says fully automated cyberattacks have arrived
- Google finds fully autonomous exploit pipelines in the wild
- Google Reports First Autonomous AI Attack That Finds and Exploits Flaws
- Five Agent Vulnerabilities in Five Months Expose Structural Flaw in MCP
- Google Cloud ships new security tools for autonomous AI agents