Google Says It Found an AI Framework That Writes and Runs Exploits From Software Patches, With No Human in the Loop

Google said it found exploit_agent, an AI framework that compares software patches, writes exploits and runs them without a human in the loop, GTIG reported.

Share this article

Featured image: a building at the Googleplex in Mountain View, California, in a July 2016 file photo. Photo: Asoundd, Wikimedia Commons, CC BY-SA 4.0. Cropped to 16:9.

Google Threat Intelligence Group said Monday, Nov. 16, 2026, that it had, for the first time, identified threat actors using large language model frameworks to fully automate both the discovery of software flaws and the launching of exploits “without a human in the loop.”

Advertisement

At the center of that finding is a multi-agent system, which Google called exploit_agent, that the group said it found on an exposed staging server. According to the GTIG report, written with Mandiant, the framework was set up to compare popular web-application code before and after patch announcements, identify possible vulnerabilities, draft exploits and run them automatically.

2025 zero-days
8 a month
Average, as reported by GTIG
2026 zero-days
11 a month
Average before an August spike, GTIG said
August 2026
22 zero-days
The spike GTIG reported for that month
August vs. 2025
+175%
TBR calculation from GTIG’s 22 and 8 figures

What GTIG said it found on the exposed server

Google said the server hosted victim data that had been taken from targeted systems, along with scripts that could be used after a compromise, including tools for internal reconnaissance, brute-forcing, installing webshells and removing other attackers’ webshells. Victim telemetry on the server, GTIG said, “suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector.”

A configuration file on the server included Russian-language comments that, in Google’s translation, referred to “LLM API keys for auto-exploitation (exploit_agent pilot).” The report did not, in that section, attribute exploit_agent to a named group or a country. The comments are evidence of the language used in the file, not, by themselves, a formal attribution.

GTIG placed the finding in a broader shift: adversaries, it said, are moving from typing prompts into a chatbot toward multi-agent frameworks that can chain together steps of an intrusion at machine speed. The report said overall “disclosures and exploits nearly doubled between January and August 2026,” while also warning that a wave of claimed, non-working, LLM-generated proof-of-concept code is creating extra verification work for defenders.

The Boston Report is describing this at a high level from Google’s published account. This article does not reproduce exploit code or attack steps.

What the data shows

In a September 2026 review of vulnerability statistics, GTIG said large language models are changing “not only the speed of vulnerability discovery and exploitation, but also the types and risk profiles of the flaws being uncovered.” Zero-day exploits, the group said, rose “only marginally” from a 2025 average of 8 per month to 11 per month in 2026, “before spiking to 22 in the month of August.”

Those are Google’s figures, not an independent count by The Boston Report. A TBR calculation from the same numbers shows the August 2026 total was 14 exploits higher than the 2025 monthly average, a 175% increase, and 11 exploits higher than the 2026 average Google cited for the period before that spike, a 100% increase. Google did not publish, in the accompanying text, a month-by-month table of n-day exploits that would allow a matching chart of those flaws.

Bar chart of zero-day exploits per month as reported by Google: a 2025 average of 8, a 2026 average of 11 before an August spike, and 22 in August 2026
Source: Google Threat Intelligence Group, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold,” Nov. 16, 2026. Percentage changes in the text are Boston Report calculations from those published figures. Chart: The Boston Report.

The report also cited outside research on speed. In late August 2026, it said, two flaws in PaperCut print-management servers were exploited as zero-days; researchers and the company Huntress generated working exploits within a day of disclosure, and GreyNoise reported a campaign four days after disclosure against “hundreds of organizations.” In a separate case, Google said official patches for two PHP vulnerabilities were released on Oct. 21, 2026, after a proof-of-concept video and early patched code had already appeared, and that “hundreds of claimed but non-functional” proof-of-concept samples, “likely generated with the help of LLMs,” followed in the first days after the official patch.

Other frameworks Google said it is seeing

exploit_agent is not the only system named in the report. GTIG said it has seen open-source offensive frameworks including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI used or tested in the wild. It said it observed CAIRN in intrusion activity that included a target in the U.S. aerospace sector, and that between late August and early October 2026 it tracked a customized platform built on PentAGI used for reconnaissance and vulnerability scanning against organizations in Asia.

Advertisement

The group also said it had seen financially motivated actors and “suspected espionage threat actors with a PRC nexus” experimenting with such tools, including evidence that a PRC-nexus group was developing its own automated attack framework. Those are GTIG’s assessments. A SecurityWeek account of the same report, published Nov. 17, summarized the same findings.

The report is the latest in a week of AI-security news. California has sued OpenAI over autonomous-agent incidents, and Google in mid-November released Gemini 4 Hydrogen, a model it has marketed for agent work, two days after opening Gemini 4 Carbon more widely to paying developers. In the same threat report, Google said Carbon includes defenses meant to block requests to develop “functional, weaponized exploits or malware tooling.” Apple, also on Monday, tightened Mac privacy settings for AI agents that request broad access to a computer’s files and apps.

What Google said it is doing in response

Google said it is using model-level safeguards, threat intelligence and “targeted containment protocols,” and that it limits or disables access for accounts it finds are misusing its systems. The company also pointed to defensive projects, including Big Sleep, an agent from Google DeepMind and Project Zero that searches for unknown software flaws, and CodeMender, which it said is used to find and fix vulnerabilities.

GTIG’s forecast was blunt. As the tools and the underlying models improve, the group said, it expects “not only reduced human involvement in these operations but also increased stealth and sophistication of the activity.”

Sources

Read next in Artificial Intelligence

Get The Boston Report Briefing

Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.

Free. We never sell your address. Privacy policy

Get The Boston Report Briefing
Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.
Sign up