Google’s threat researchers have delivered the starkest warning of the year for anyone who runs software: the hacker no longer needs to be in the loop. In its latest AI Threat Tracker, published on November 16, 2026, the Google Threat Intelligence Group (GTIG) and Mandiant say they have observed attackers using large language model frameworks to fully automate the discovery and exploitation of software vulnerabilities — for the first time. The report’s blunt verdict, in the words of GTIG chief analyst John Hultquist, is that “the era of autonomous attacks has arrived.”
The finding matters because it changes the arithmetic of patching. Until now, a security fix gave defenders a head start: someone still had to read the patch, work out what it fixed, write an exploit and launch it. Google says that chain is now running on its own, at machine speed, and the window between a patch going public and an attack arriving is collapsing.
What Google found: exploit_agent turns patches into exploits
The evidence sits on an exposed staging server GTIG discovered, hosting a multi-agent framework the researchers call exploit_agent. The system plugs into both frontier and locally hosted language models and runs a continuous loop: it watches popular web applications for new patches, compares the code before and after each fix, works out which flaw was closed, drafts an exploit for it and then runs that exploit automatically — with no human approving any step.
Google printed one example of the framework at work: it was inspecting access-control code in Spectra, the Ultimate Addons for Gutenberg WordPress plugin at version 2.19.25, a plugin with more than a million active installations. Google is careful about the limit here — it does not say the agent found a flaw in Spectra or exploited it, only that the framework was reading the code. What it does say is that the same server held stolen victim data and scripts for the steps after a break-in, including scanning inside networks, guessing passwords at high speed and installing web backdoors. Victim telemetry on the server pointed at French media, entertainment and education websites.
The report is equally careful about attribution. Google names no group or country behind exploit_agent, and notes that Russian comments found in a configuration file are not, on their own, enough to attribute it. The framework’s existence is the story; who ran it remains unknown.

Stage 5 of 7: where autonomous attacks stand now
To make the shift concrete, Google sorts attacker use of AI into a seven-stage timeline, published as Figure 2 of the report under the title “Evolution of threat actor use of LLMs.” The first four stages cover the road already travelled: prompting chatbots, embedding AI features in malware, adding automation to operations, and experimenting with autonomous frameworks. Today, Google says, sits at Stage 5: “Multiple attack lifecycle stages conducted by LLM-driven tooling in-the-wild,” with minimal human involvement.
The next two stages are forecasts, not observations, and Google labels them as such. Stage 6, projected for the next one to three months, is multi-agent-driven, unsophisticated attacks with no human in the loop. Stage 7, projected for three to six months out, is sophisticated, fully autonomous or agentic attacks. The distinction matters: fully automated exploitation of known flaws has been observed; fully autonomous intrusions end to end have not — yet. The direction of travel, though, is unambiguous, and the forecast window is measured in months, not years.
The open-source toolkits behind the shift
exploit_agent did not appear from nowhere. The report describes a broader move by attackers toward open-source, multi-agent penetration-testing frameworks that were built for legitimate security testing and are now being abused in live intrusions. GTIG names five: ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. It says it has independently observed CAIRN and PentAGI deployed in the wild and has seen experimentation with CyberStrikeAI and HexStrike AI.
One case shows the scale these tools enable. Between late August and early October 2026, GTIG tracked a threat actor using a custom attack-orchestration platform built on the open-source version of PentAGI to run systematic reconnaissance and vulnerability scanning against a wide range of organizations in Asia. Separately, the FBI issued a FLASH advisory on November 13 warning that criminal actors have weaponized the ARTEX framework since at least August 2026, using it to automate phases of an intrusion from reconnaissance through data exfiltration.
The pattern is consistent: the frameworks chain known weaknesses at machine speed, so a defender’s patching window is now competing with an automated pipeline rather than a human analyst’s working hours. The industry is responding in kind: NVIDIA has handed its OpenShell guardrails for autonomous AI agents to the CNCF, putting safety rails around the same agent frameworks attackers are abusing.
The numbers: vulnerability disclosures nearly doubled in 2026
The report also puts numbers behind the pressure. Vulnerability disclosures nearly doubled between January and August 2026, rising from 5,045 in January to 10,477 in July and 10,740 in August, according to GTIG’s September analysis. Exploitation rose with it, from 10.5 cases a month to 18 a month. Zero-day exploits — flaws attacked before a fix exists — climbed from a 2025 average of 8 a month to 11 a month in 2026, before spiking to 22 in August.
Two PHP flaws, tracked as CVE-2026-109618 and CVE-2026-109619, illustrate the new dynamic: patches for both landed on October 21, 2026, and the report’s differential-analysis pipeline is exactly the kind of tooling that can turn a published patch into a working exploit within hours. That is why the same week brought such a heavy Patch Tuesday — Microsoft’s November update fixed 1,094 flaws, including an Outlook zero-day already under attack — and why flaws already on CISA’s exploited list, from Zyxel gear to WordPress core, need patching the moment a fix appears rather than at the next convenient maintenance window.
The same report’s second warning: coding agents as a backdoor
The Threat Tracker contains a second finding that should worry development teams just as much. GTIG says a suspected North Korean cluster it calls MIDNIGHT NEPTUNE edited .claude/settings.json files inside internal code repositories, so that launching Claude Code automatically downloaded a backdoor GTIG names SOMBERMEME. The intrusions, which targeted cryptocurrency organizations in North America and Europe, began with fake recruiter personas and ClickFix lures.
The technique is a supply-chain pivot: instead of attacking the coding agent itself, the attackers tampered with the agent’s committed configuration. It is a reminder that as AI coding assistants become part of the build pipeline, their settings files deserve the same scrutiny as any other executable code — and the same code review before they are merged.
What defenders should do now
Google’s report is a warning, but it also sketches the response. The company points to its own Google AI Threat Defense architecture and BeyondZero, SecOps rule packs for detecting ARTEX activity, and its CodeMender and Big Sleep agents, which use AI on the defensive side to find vulnerabilities and draft fixes. The broader lesson, though, applies to every team:
- Patch faster, and treat patch day as exploit day. With differential analysis automated, the gap between a fix being published and an exploit existing can now be hours. Prioritize internet-facing software and anything already on CISA’s exploited list.
- Use virtual patching where real patching lags. If a fix cannot be applied immediately, a web application firewall rule or vendor mitigation buys time against automated scanning.
- Audit your AI agent configurations. Review settings files, hooks and plugins for coding agents and other autonomous tools the way you would review code — before they run, not after.
- Watch for the frameworks, not just the flaws. Detection rules for ARTEX, PentAGI and similar tooling can flag an intrusion in progress even before a specific vulnerability is identified.
- Fight agents with agents. The same week Google warned about autonomous attackers, it also showed defenders using AI agents to find and fix flaws faster, including AI agents that attack your code and then patch it — the defensive use of the technology is the other half of this race.
The era Google describes is not a distant threat. A framework that reads a patch, writes the exploit and fires it off without a human has already been found running in the wild. The teams that adapt their patching speed and their monitoring to machine-speed attackers are the ones that will stay ahead of it.

Frequently asked questions
What is exploit_agent?
exploit_agent is a multi-agent AI framework Google’s Threat Intelligence Group found on an exposed staging server. It continuously compares software code before and after security patches, identifies the flaw each patch fixed, drafts an exploit and executes it automatically, using both frontier and locally hosted language models.
Is this the first fully autonomous cyberattack?
Not quite. Google says this is the first time it has observed attackers fully automating vulnerability discovery and exploitation with no human in the loop. A complete, fully autonomous intrusion from start to finish has not been observed yet — Google forecasts that stage within three to six months.
What does Stage 5 mean in Google’s timeline?
Stage 5 of Google’s seven-stage timeline means multiple stages of an attack are being run by LLM-driven tooling in the wild, with minimal human involvement. Stages 6 and 7 — attacks with no human in the loop, then sophisticated fully autonomous attacks — are forecasts for the next one to six months.
Which tools are attackers abusing?
GTIG names open-source multi-agent penetration-testing frameworks including ARTEX, CAIRN, CyberStrikeAI, HexStrike AI and PentAGI. The FBI separately warned on November 13 that ARTEX has been weaponized by criminal actors since at least August 2026.
What should small teams do first?
Patch internet-facing software immediately when fixes land, enable automatic updates where possible, and audit the configuration files of any AI coding assistants your team uses. Those three steps cover the two attack paths in Google’s report: automated patch-to-exploit pipelines and tampered agent settings.

