Autonomous Cyber Attacks Are Here: What Google's GTIG Found

By Eoin Villarin, 2026-11-17

Google Threat Intelligence Group reports the first AI framework that turned security patches into working exploits with no human in the loop. Here is what it found and how to defend your site.

What are autonomous cyber attacks and what did Google find?

Autonomous cyber attacks use AI agents that find, exploit and act on vulnerabilities without a human approving each step. In its November 2026 AI Threat Tracker, Google Threat Intelligence Group found exploit_agent, a framework that compared software patches, wrote exploits and ran them automatically against unpatched websites, mostly French media, entertainment and education sites. GTIG calls it the first fully automated discovery and exploitation it has seen.

Key takeaways

  • Google GTIG found exploit_agent, an AI framework that turned patches into exploits with no human in the loop.
  • Attackers now favor n-day bugs because a public patch shows the AI exactly what to exploit.
  • Open-source agent tools like CAIRN, PentAGI and HexStrike AI were used or tested in real intrusions.
  • LLMJacking cases cost victims over $100,000 in a day and up to $1 million in the worst cases.
  • Fast patching, MFA or passkeys, scoped API keys, rate limits and log alerts block most of this activity.

"Cyber attack" is trending because Google says autonomous cyber attacks are no longer a forecast. In its latest AI Threat Tracker, published November 16 and 17, 2026, Google Threat Intelligence Group (GTIG) reports the first case it has seen of attackers using an AI framework to find vulnerabilities, write exploits and run them with no human in the loop. The tool, which GTIG calls exploit_agent, compared software before and after security patches, worked out what each patch fixed and attacked sites that had not updated yet. Below is what GTIG actually found, which threat actors and AI tools it names, what Google did, and a practical defense checklist for small businesses, site owners and developers.

GTIG AI Threat Tracker at a glance

ItemDetails
ReportGTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold (Google Cloud blog), plus the summary "The era of autonomous attacks has arrived" on blog.google
PublishedNovember 16, 2026 (blog.google) and November 17, 2026 (Google Cloud blog)
WhoGoogle Threat Intelligence Group with Mandiant
Headline findingexploit_agent, a multi-agent framework found on an exposed staging server, that diffs patches, drafts exploits and runs them automatically
Main targets of exploit_agentFrench websites in media, entertainment and education, based on victim telemetry
Exploit trendZero-day exploits went from an average of 8 per month in 2025 to 11 per month in 2026, then 22 in August
LLMJacking lossesSingle-day fraudulent AI charges above $100,000 in some cases, up to $1 million cumulative in the worst
Google responseDisabled abusive accounts and assets, updated Gemini classifiers and models, detection rules in Google SecOps, Mandiant hardening guide, Google AI Threat Defense

What are agentic or autonomous cyber attacks?

Until recently, most AI use by attackers meant a person typing prompts into a chatbot. An agentic attack is different. The attacker sets up a framework where one or more AI agents plan steps, call real tools (scanners, exploit scripts, shells), read the results and decide what to do next on their own. GTIG describes this as a move from manual prompting to autonomous multi-agent frameworks and automated exploitation pipelines.

An autonomous attack is the far end of that scale: the human sets the goal and the infrastructure, then the software runs discovery, exploitation and parts of post-compromise work without someone approving each step. GTIG says that in the past quarter it saw this happen for vulnerability discovery and exploitation for the first time. In its previous tracker, about two months earlier, it had described only partially automated scanning and exploitation chains.

Google data center buildings and cooling towers in The Dalles, Oregon, with hills behind
A Google data center in The Dalles, Oregon. GTIG and Mandiant work from Google's telemetry and incident response cases. Photo: Tony Webster, CC BY 2.0, via Wikimedia Commons

What exactly did GTIG find?

exploit_agent: patch to exploit with no human

GTIG found an exposed staging server running a multi-agent framework that connects to several frontier and locally hosted LLMs. GTIG does not name the specific models. The framework continuously compares popular web application code before and after each patch, flags the likely bug, drafts an exploit and executes it automatically. The same server held data stolen from victims and scripts for internal reconnaissance, brute-forcing, installing webshells and removing webshells left by other attackers.

Configuration comments on the server were written in Russian, according to the report, but Google does not attribute the operator to a named group or country. GTIG also did not publish a success rate or a count of compromised sites, a gap that critics such as SnowRock pointed out.

Why patches are now the attack map

The most useful point for site owners: GTIG thinks attackers may get more value pointing LLMs at n-day vulnerabilities (bugs already fixed in a public patch) than hunting zero-days. A patch shows exactly what changed, so an AI can work backwards to the flaw. GTIG Chief Analyst John Hultquist told SecurityBrief Asia that "Enterprises are going to be under enormous pressure to speed up their patch cycles". In plain terms, the gap between a plugin update and the first automated attack on unpatched sites is shrinking.

PaperCut: exploits within a day

GTIG cites two chained PaperCut print server flaws from late August 2026, CVE-2026-81578 (authentication bypass) and CVE-2026-82078 (unsafe dynamic class loading), which together allowed unauthenticated remote code execution. Researchers and Huntress produced working exploits within a day of disclosure. GreyNoise then tracked a campaign four days after disclosure that hit hundreds of organizations; in one case an initial foothold reached domain administrator within minutes.

Open-source agentic pentest tools used in real intrusions

GTIG says attackers are abusing open-source agentic pentesting frameworks to automate parts of live intrusions:

  • CAIRN, a multi-agent framework, was observed in intrusion activity against targets including the US aerospace sector.
  • PentAGI: from late August to early October 2026, an actor ran a customized platform built on it for reconnaissance and vulnerability scanning against a wide range of organizations in Asia.
  • ARTEX, CyberStrikeAI and HexStrike AI are also named. HexStrike AI is an MCP (Model Context Protocol) harness that exposes more than 150 offensive tools to LLM clients.

The good news in the report: these frameworks currently trade stealth for speed. They generate unusual traffic, recognizable identifiers and code artifacts that defenders can detect.

LLMJacking: stolen AI access sold for profit

Mandiant responded to several cases where attackers stole cloud or AI credentials (often from public GitHub repositories) and ran up AI and GPU bills. GTIG mentions CVE-2026-40217, a sandbox escape in the LiteLLM proxy that turns limited "Virtual Keys" into full admin credentials.

State-linked actors and AI coding tools

  • CALANQUE ION (formerly APT42, Iran) used Gemini for reconnaissance and targeted social engineering.
  • A suspected North Korean cluster GTIG calls MIDNIGHT NEPTUNE used fake recruiter profiles with AI images, ClickFix lures, and changed .claude/settings.json files in internal repositories so that starting Claude Code downloaded a backdoor called SOMBERMEME.
  • Influence operation vendors used the Gemini API and other providers to automate fake personas at scale.

What did Google do about it?

According to the report, Gemini's safety systems triggered on parts of this activity, Google disabled the accounts and assets tied to it, and Google DeepMind used the findings to strengthen classifiers and models so they refuse similar requests. Google SecOps shipped detection rulepacks for ARTEX, and Mandiant released a hardening guide aimed at LLM-driven offensive tooling. On the defensive side, Google points to Big Sleep and CodeMender for finding and patching bugs, and to Google AI Threat Defense, which combines Gemini, Wiz, CodeMender and Mandiant. The blog.google summary puts the core argument simply: "Defending against machine-speed threats at human speed isn't feasible."

A fair caveat: the warning arrived alongside a product pitch, and the evidence for full autonomy is one framework on one server. Separately, some outlets bundled the GTIG news with the Optimizely script hijack; that is a different incident, covered in its own post.

Why does this matter for small businesses and site owners?

Small sites were never too small to attack, because most attacks are automated scans for known holes. What changes now is speed and volume. If an AI pipeline can turn a WordPress plugin patch into a working exploit in hours, the old habit of updating "when I get around to it" becomes the main risk. exploit_agent targeted ordinary media, entertainment and education websites, not banks or governments.

Rows of server towers on metal shelves lit with blue LED strips in a data center
Automated attacks run from rented servers around the clock, so patch and alert windows need to shrink too. Photo: BalticServers.com, CC BY-SA 3.0, via Wikimedia Commons

AI cyber attack defense checklist

None of this needs an enterprise budget. This is the order I work through for client sites.

1. Cut your patch time to days, not weeks

  1. List every component: CMS core, plugins, themes, npm and Composer packages, server OS, VPN and self-hosted tools.
  2. Turn on automatic minor and security updates where safe, and subscribe to vendor security advisories.
  3. Set a target: critical internet-facing fixes within 48 hours, everything else within 7 days.
  4. Remove plugins and services you no longer use. Unused code still gets diffed and exploited.

If you do not have time for this, a website maintenance plan exists exactly for this job.

2. MFA everywhere, passkeys where possible

Turn on MFA for hosting, registrar, CMS admins, email, GitHub and every cloud or AI console. Prefer passkeys or hardware keys over SMS, because phishing-resistant login makes stolen passwords far less useful.

3. Least privilege API keys and secrets hygiene

  • Never commit keys. LLMJacking cases often began with credentials found in public GitHub repositories. Enable secret scanning and push protection.
  • Give each app its own key with the smallest scope, and set hard spending limits and quota alerts on AI and cloud accounts.
  • Rotate keys on a schedule and immediately after any staff change or leak.
  • If you run LiteLLM or a similar AI proxy, patch it and do not expose its admin interface to the internet.

4. Rate limiting and bot protection

Agentic frameworks are noisy: many requests, fast retries, brute force. Rate limit login, password reset, search and API endpoints, and put a WAF or bot protection layer in front of the site. A basic nginx example:

limit_req_zone $binary_remote_addr zone=login:10m rate=5r/m;

server {
    location /api/login {
        limit_req zone=login burst=5 nodelay;
        limit_req_status 429;
        proxy_pass http://app_upstream;
    }
}

5. Logging and alerting you will actually read

  • Send web server, auth and admin logs somewhere off the server, so an attacker cannot wipe them.
  • Alert on new admin users, plugin installs, file changes in upload folders and spikes of 404 or 500 errors.
  • Watch AI and cloud billing daily. A sudden spend jump is often the first sign of LLMJacking.

Since exploit_agent installed webshells, a quick check for PHP files where none should exist is worth running on any WordPress server:

find /var/www/html/wp-content/uploads -type f -name "*.php" -mtime -14 -print
Small office network rack with patch panels, two Ethernet switches and a router with neatly bundled grey cables
A typical small business network rack. Patching and logging the gear in here matters as much as the website. Photo: Dsimic, CC BY-SA 4.0, via Wikimedia Commons

6. Secure AI agents, coding assistants and MCP tool access

The MIDNIGHT NEPTUNE case shows that AI tool configuration files are now part of your supply chain. For developer teams:

  • Treat .claude/, .cursor/, .vscode/ and MCP config files like CI pipelines: require code review for any change.
  • Only install MCP servers from sources you trust, pin versions, and give each one the narrowest permissions and tokens it needs.
  • Run agents with no production credentials by default, and keep "auto approve" off for shell commands and file writes outside the project.

A CODEOWNERS file makes the review rule enforceable on GitHub:

/.claude/ @acme/security
/.mcp.json @acme/security
/.github/workflows/ @acme/security

7. Prompt injection defenses for sites that run chatbots

The GTIG report does not cover prompt injection directly, but if your site runs a support bot or AI agent with tools, the same agentic logic applies to you as a target. Assume any text the model reads (user messages, web pages, uploaded files) can contain instructions. Enforce permissions in your code, not in the system prompt:

const ALLOWED_TOOLS = new Set(["search_docs", "get_order_status"]);

function runTool(call, session) {
  if (!ALLOWED_TOOLS.has(call.name)) {
    throw new Error("Tool not allowed");
  }
  if (call.name === "get_order_status" && !session.orderIds.includes(call.args.orderId)) {
    throw new Error("Order does not belong to this user");
  }
  return tools[call.name](call.args);
}
  • Give the bot read-only tools by default, and require a human click for refunds, emails or account changes.
  • Never put API keys or internal URLs in the system prompt.
  • Rate limit the chat endpoint and log every tool call.

Quick priority table

WhoDo this weekDo this month
Small business ownerUpdate CMS and plugins, MFA on hosting, email and registrarMaintenance plan, off-site backups, billing alerts
Site owner or agencyRate limit logins, scan uploads for PHP files, remove unused pluginsWAF or bot protection, central logging with alerts
Developer teamSecret scanning, rotate exposed keys, review AI tool config filesCODEOWNERS for agent configs, least privilege MCP tokens, chatbot tool allowlists

If you want a second pair of eyes on where your site stands, a website audit covers outdated components, exposed admin panels and missing security headers.

Sources

FAQ

What is exploit_agent?

exploit_agent is the name GTIG gave to a multi-agent AI framework found on an exposed staging server. It connects to frontier and local LLMs, compares web application code before and after patches, drafts exploits and runs them without human approval.

Is this the first fully autonomous cyber attack?

GTIG says it is the first time it has identified attackers using LLM frameworks to fully automate vulnerability discovery and exploitation with no human in the loop. A person still built and launched the system, and Google did not publish how many sites it compromised, so "first autonomous discovery and exploitation pipeline" is more accurate than "AI hacked on its own."

Which AI models did attackers misuse?

GTIG does not name the models behind exploit_agent. It names Gemini as used by the Iranian group CALANQUE ION and the Gemini API in influence operation platforms. A North Korean cluster abused Claude Code configuration files to deliver a backdoor, and attackers used open-source agent frameworks like CAIRN, PentAGI and HexStrike AI.

Are small business websites at risk from AI cyber attacks?

Yes, mainly through speed. AI can turn a public patch into an exploit quickly, so sites running outdated plugins, themes or server software are the easiest targets. exploit_agent went after ordinary media and education sites, not only large companies.

What is LLMJacking?

LLMJacking is when attackers steal cloud or AI credentials and use the victim's account to run AI models or GPUs, or resell the access. Mandiant saw single-day charges above $100,000 and cumulative abuse up to $1 million in the worst cases. Spending limits, key scoping and billing alerts are the main defenses.

Related reading