Hot AI News

Google found an AI tool that turns patches into exploits on its own

Google's Threat Intelligence Group says it has for the first time seen attackers fully automate vulnerability discovery and exploitation with an AI framework, with no human in the loop.

Google found an AI tool that turns patches into exploits on its own
Photo: OutreachPete / Wikimedia Commons, CC BY 2.0

Google’s Threat Intelligence Group (GTIG) says it has for the first time seen attackers use an AI framework to find and exploit vulnerabilities with no human in the loop, as first flagged by Techmeme from SiliconANGLE’s coverage. The tool, called exploit_agent, turns fresh security patches into working exploits.

The finding is in GTIG’s quarterly AI threat report, “Agentic Threat Activity Crosses a Significant Threshold”, published November 16, 2026 by GTIG and Mandiant.

A Google sign at 1600 Amphitheatre, the Googleplex campus in Mountain View, with building direction arrows.
Photo: Runner1928 / Wikimedia Commons, CC BY-SA 3.0

What exploit_agent does once a patch ships

GTIG says it found exploit_agent on an exposed staging server. It describes the tool as a multi-agent framework for vulnerability research and weaponization that connects to several frontier and locally hosted language models.

According to the report, the framework continuously compares popular web applications before and after patch announcements. It analyzes the differences to spot a likely flaw, drafts an exploit and runs it automatically. One figure in the report shows it analyzing access control in version 2.19.25 of the Spectra (Ultimate Addons for Gutenberg) WordPress plugin.

The same server held scripts for internal reconnaissance, brute-forcing, and installing webshells, plus a script that removes other attackers’ webshells. Victim telemetry suggests the operator targeted French websites in media, entertainment and education. Comments in an environment file, translated from Russian, show API keys set up for “auto-exploitation (exploit_agent pilot).”

Advertisements

Why patched flaws are the easier target

GTIG’s reasoning is that a patch hands attackers a finite body of source material. Pointing language models at published fixes, known as n-day vulnerabilities, may be more efficient than hunting for unknown flaws. The group says models help attackers compare code versions, release notes and proof-of-concept code, which shrinks the window defenders have to patch.

The report also gives defenders some room. Exploits that language models claim will work often do not, which adds verification work. And GTIG says the frameworks compress attack timelines but leave noisy forensic traffic, recognizable identifiers and code-level traces, which raises the risk of detection. It expects that to change as tools and models improve, with less human involvement and more stealth.

Zero-day exploits ran at 22 in August 2026, against an average of 8 a month in 2025.

The numbers behind a shrinking patch window

GTIG puts the pressure in figures. Disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits averaged 8 per month in 2025 and 11 per month in 2026, then hit 22 in August alone.

The report ties this to concrete cases. Two chained flaws in PaperCut print management servers were exploited as zero-days in late August (CVE-2026-81578 and CVE-2026-82078). GTIG says working exploits were generated within a day of disclosure. SiliconANGLE adds that researchers and Huntress Labs used language models to do so, and that a campaign tracked by GreyNoise began four days after disclosure and hit hundreds of organizations, with one foothold reaching domain administrator within minutes.

This is a change from GTIG’s previous edition. About two months earlier, according to SiliconANGLE, the group said it had not observed fully autonomous attack pipelines in the wild. Our coverage of Kindo’s $30M round shows the other side of the same trend: companies are now paying to control what AI agents can touch.

Advertisements

What the report does not settle

The evidence is one exposed staging server and the telemetry on it. The report does not say how many victims were actually compromised by exploit_agent, who runs it, or which model providers it used. Attribution of the operator rests on a Russian-language comment, not a named actor. All findings are Google’s own and have not been independently verified.

GTIG also describes wider agentic activity. Open-source pentesting frameworks such as CAIRN and PentAGI were seen in real intrusions, and CyberStrikeAI and HexStrike AI were seen in experimentation. Suspected PRC-nexus espionage actors and financially motivated groups were among the users.

For defenders, the practical point is speed: the gap between a patch landing and a working exploit is shrinking, so patch cadence matters more. Google says it disabled assets linked to Gemini misuse, shared findings with Google DeepMind to strengthen its classifiers, and released detection rules for the ARTEX framework. Mandiant also published a hardening guide.

Sign Up for the Daily AI Pulse

One email a day. All the stories that matter.

Daniel ReedAI Research, Safety & Ethics Analyst

Daniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…

Leave a Reply

Your email address will not be published. Required fields are marked *

You may also like