Google Sees First No-Human AI Exploit Tool; Zero-Days Hit 22

Google’s Threat Intelligence Group (GTIG) said in its AI Threat Tracker report, published on 16 November, that in the past quarter it identified threat actors using LLM frameworks “to fully automate vulnerability discovery and exploitation without a human in the loop for the first time”. The framework, which GTIG dubbed exploit_agent, sat on an exposed staging server and “integrates with multiple frontier and local LLMs”.

The same report says zero-day exploits rose from a 2025 average of 8 per month to 11 per month in 2026, “before spiking to 22 in the month of August”. August’s 22 is 2.75 times the 2025 monthly average, by our arithmetic.

Business Pill · AI Agents Explained

A short Business Engineer explainer on AI agents, software that carries out multi-step tasks. It teaches the general idea only and does not discuss Google’s report or cyber attacks.

The key insight: As we read the report, the step GTIG calls new is not the exploit itself but the missing human: a framework on one server compared code before and after patch announcements, drafted exploits and ran them, while zero-days reached 22 in August, 2.75 times the 2025 monthly average of 8, by our arithmetic.

What GTIG Found on the Server

GTIG describes exploit_agent as a “multi-agent vulnerability research and weaponization framework”. The report says it “compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention”.

The framework was set up for “continuous differential analysis of popular web application software”. One figure in the report shows it inspecting access-control code in the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25.

GTIG says it also found data exfiltrated from targeted victims and scripts for internal reconnaissance, brute-forcing, installing webshells and “removing other attackers’ webshells”. Victim telemetry, the report says, suggests the operator was interested in French websites in media and entertainment, and in the education sector.

The report reproduces comments in Russian from a configuration file, which it translates as “LLM API keys for auto-exploitation (exploit_agent pilot)”. GTIG does not name the operator or say which models the framework called.

Zero-day vulnerabilities exploited per month as counted by Google Threat Intelligence Group: a 2025 average of
Zero-day vulnerabilities exploited per month as counted by Google Threat Intelligence Group: a 2025 average of 8, a 2026 average of 11 and 22 in August 2026. Source: GTIG, 16 November 2026; 2.75x is our arithmetic.

The Patch Window

GTIG’s case is that chaining these steps lets attackers “dramatically reduce the window between patch disclosure and weaponized exploitation”. It says LLMs are “shrinking the patch window for organizations”.

The report cites outside research as corroboration. For two PaperCut print-server flaws exploited as zero-days in late August 2026, CVE-2026-81578 and CVE-2026-82078, it says security researchers and Huntress used LLMs to generate working exploits “within a day of disclosure”. It says GreyNoise reported a campaign four days after disclosure against “hundreds of organizations”, and that the attacker moved from a foothold to domain administrator “within minutes”.

It also cites two PHP flaws, CVE-2026-109618 and CVE-2026-109619, patched on 21 October 2026. In the first days after the patch, GTIG says, hundreds of claimed but non-functional proofs of concept, likely generated with the help of LLMs, were published, “creating additional verification work for defenders” in its words.

The Frameworks and the Bills

Beyond exploit_agent, GTIG names open-source multi-agent frameworks in attackers’ hands: ARTEX, CAIRN, CyberStrikeAI and PentAGI. It says CAIRN was used in intrusions against targets including the US aerospace sector, and that a PRC-nexus group was developing its own automated attack framework.

The report adds a cost line for companies. In LLMJacking cases handled by Mandiant, where attackers run up a victim’s cloud and AI bills, GTIG says single-day fraudulent charges exceeded $100,000, and in the most severe cases cumulative abuse reached $1 million before remediation.

Our Mastercard piece covered the CEO’s $16 trillion cyber-damage projection; our GPT-6 Cyber piece compared OpenAI’s and Anthropic’s own cyber scores, and our Mythos tiers piece set out how Anthropic gates defenders’ access.

What Google Says It Is Doing

GTIG lists its own responses. Google Security Operations “has deployed specialized behavioral and volumetric rulepacks to detect use of ARTEX”, and Mandiant released a hardening guide. On the defence side Google points to Big Sleep and CodeMender, its AI agents for finding and fixing flaws, and to Google AI Threat Defense, which combines Gemini with Wiz and CodeMender.

The report also says Google recently introduced Gemini 4 Carbon, “our strongest model with state-of-the-art cybersecurity and long horizon capabilities”, built with layered defences against cyber misuse. Our chip supply-chain explainer covers the other end of the same risk.

The Structural Read

As we read the cost side, the bill shows up in two places GTIG names: the time defenders have between a patch and an attack, which it says LLMs are shrinking, and cloud and AI charges run up with stolen keys, which it puts above $100,000 in a single day in some cases.

As we read Google’s position, the same report that documents the attacks lists Google’s own defensive products, from Big Sleep and CodeMender to Google AI Threat Defense; the report is both threat research and the context for those products.

Google Threat Intelligence Group, 16 November 2026

“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.”

Three Implications

FOR SECURITY TEAMS GTIG says LLM pipelines can dramatically reduce the window between patch disclosure and weaponized exploitation; the PaperCut flaws it cites had working exploits within a day of disclosure.

FOR CLOUD AND AI BUDGETS In LLMJacking cases, GTIG says single-day fraudulent charges exceeded $100,000 and the most severe cases reached $1 million before remediation.

FOR THE COUNT Zero-days rose from a 2025 average of 8 a month to 11 in 2026 and 22 in August, per GTIG; 22 is 2.75 times 8, by our arithmetic.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The analysis says: “The surface, not the model, governs autonomy.”

As we read it, exploit_agent fits that frame from the attack side: GTIG describes a harness wired to multiple frontier and local LLMs, and the autonomy sits in that harness. The framework itself discusses usage data from OpenAI and Anthropic, not cyber attacks.

What Is Not Established

GTIG does not say how many victims exploit_agent reached, how long it took from a given patch to a working exploit, which frontier models it used, or who ran it. The Russian-language comments are GTIG’s finding; the report does not attribute the framework to a named group or country.

The zero-day counts are GTIG’s, from its September analysis of January 2025 to August 2026. The report does not tie the August spike to exploit_agent, and the PaperCut and PHP timelines come from the outside research it cites. Google was not contacted for this piece.

This Week In The AI Supercycle

AI Lost Its Central Banker, the Lab Oligopoly, and the Future of Open Source

This week’s issue: the Nvidia leadership change, read as the AI build-out losing its central banker: who now finances the GPU landlords, backstops the data centres and sponsors open models, and why two labs are becoming the bank.

Read the full issue →

Business Engineer Framework

The Agentic Harness War

The Business Engineer analysis of where value pools in the agentic AI stack: the harness around the model.

Read the Map of AI →

The Bottom Line

Google’s threat team says it has now seen an AI framework take a patch, find the flaw, write the exploit and run it with no human in the loop, in a year when zero-days reached 22 in August, 2.75 times the 2025 monthly average of 8, by our arithmetic.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Every finding, quote, framework name, CVE number, dollar figure and zero-day count comes from Google Threat Intelligence Group’s AI Threat Tracker post of 16 November 2026, read in full on Google Cloud’s blog; the PaperCut and PHP timelines are outside research that GTIG cites. The 2.75 times figure is our arithmetic. Google was not contacted for this piece.

Sources: Google Threat Intelligence Group: GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold (16 Nov 2026)

Scroll to Top
FourWeekMBA