Google Says Autonomous AI Cyberattacks Have Arrived

On 16 November 2026, Google's Threat Intelligence Group published its Q4 2026 AI Threat Tracker. According to the report, it has found the first case of attackers using an AI framework to find software flaws and write working exploits with no human directing each step. The framework is called exploit_agent. Google expects simpler autonomous attacks within one to three months and more capable ones within three to six months. For a business that automates routine work, the practical point is that the software and accounts behind its automations now sit in a more dangerous environment. Access rights, patch timing and approval steps for payments and bank changes need a fresh look.
In short
- Google's Threat Intelligence Group reported on 16 November 2026 its first case of fully automated vulnerability discovery and exploitation.
- The framework, exploit_agent, compares patched code, drafts exploits and runs them without a human directing each step.
- Google expects simpler autonomous attacks within one to three months and more capable ones within three to six months.
- Businesses should narrow the access their automations hold and keep people in the approval step for payments and bank changes.
- Patch speed and job-offer verification matter more when automated tools hold logins to business systems.
Google's Threat Intelligence Group, working with the security firm Mandiant, published its AI Threat Tracker on 16 November 2026. The report describes what Google calls the first case of fully automated vulnerability discovery and exploitation by malicious actors. The example is a framework named exploit_agent. Google's researchers found it on an exposed staging server, a machine used by the attacker to prepare operations.
What Google found
According to the report, exploit_agent is a multi-agent system. It connects to several large language models, which are AI systems that generate text and code. It compares a web application's code before and after a software patch is released. From that comparison it works out which flaw the patch closed, drafts an exploit for the flaw, and runs the exploit. Google says the loop runs without a person supervising each step.
Google linked the operator's interest to French media, entertainment and education websites. The report also describes separate operations against organisations in the United States aerospace sector and in parts of Asia. Google did not name a country or group behind the framework. It also did not confirm that exploit_agent caused the thefts found on the server.
The same report records state-linked actors using AI in other ways. According to the report, China-linked actors have used language models for reconnaissance and research. North Korean operators have used AI to support IT worker schemes, including writing résumés and preparing for job interviews at Western companies. Iranian state-linked activity is also covered. Google's earlier 2025 reports described the same groups using AI for phishing and malware research, so the 2026 findings continue an existing pattern.
Why this matters for businesses that automate work
Most businesses do not face a rogue attacker with an autonomous tool directly. They face the same software their automations depend on: web applications, plugins, accounting and HR systems, and the cloud accounts that link them. When a vendor publishes a patch, the gap between that release and the moment customers install it has always been a window for attackers. Google's report suggests that window can shrink to machine speed.
Google's own summary puts the point plainly: defending against machine-speed threats at human speed is not feasible. That is the reason the company promotes automated defence. Automation cuts both ways. The same kind of software that lets a business post invoices, chase payments or update records can also be turned into a target if its access is too broad.
What changes in practice, at any size
For a five-person firm and for a group with thousands of staff, the first changes are similar. The steps below are a starting point, not a complete security programme.
- List every automation that holds a login, API key or token. Note which system it writes to and who can change it.
- Give each automation the narrowest access it needs. A tool that reads invoices should not also be able to change supplier bank details.
- Keep a person in the approval step for payments, new vendor accounts and changes to bank details. Verify these by phone on a number you already hold.
- Apply security patches to the web applications and plugins in your automations within days, not weeks. Confirm which version each system runs.
- Treat unexpected job offers, recruiter approaches and video calls with the same caution as any other unsolicited request for access.
The table below shows where the new pressure falls in a typical back-office workflow, and what a person should still check.
| Workflow | What the automation does | What a person still checks |
|---|---|---|
| Vendor invoices and payments | Reads invoices, matches them to orders, prepares payment runs | Any change to vendor bank details and every payment above the level you set |
| Customer website forms and portals | Pushes form entries into the CRM or billing system | Patch status of the web plugin, and access logs for unusual activity |
| Recruitment and onboarding | Schedules interviews and sends offer documents | Identity of new contacts, and any request for system access before a contract is signed |
| Reports and documents | Collects data and builds monthly reports | The source data and the final figures before they leave the company |
What automation cannot do here
Automation cannot judge whether a patch is safe to install on your systems, and it cannot decide whether a request is genuine. It also cannot replace the access review that tells you which accounts can reach which data. Google's report also notes that autonomous attacks leave traces, and that its own researchers found the exploit framework on a server that was exposed by mistake. A business still needs monitoring and someone to read the alerts.
What to watch next
Google expects unsophisticated autonomous attacks within one to three months and more sophisticated versions within three to six months, according to the report as summarised by Crypto Briefing. Watch for three things. First, whether the open-source offensive frameworks named in the report, such as CAIRN and PentAGI, spread beyond the cases Google has already seen. Second, whether vendors shorten their patch release cycles in response. Third, whether Google's own defence tools, which the company describes as combining its Gemini models with Mandiant and Wiz expertise, reach ordinary business customers. Google's report does not give a success rate for the exploits, so the speed of real-world impact remains open.
The report is a warning about where attackers are heading, not a measure of how many businesses have already been hit. Read the original report before making decisions about your own systems.
Sources
How AiStaffo would automate this
AiStaffo would start by listing every automation a client runs that holds a login, token or API key, including invoice readers, form connectors and report builders. Each one would be connected with the minimum access it needs, and changes to vendor bank details would be routed to a named person for a phone check. Routine work such as matching invoices, chasing payments and preparing reports would still run automatically, with a log of every action. The owner would still approve payments above an agreed limit, review access changes and decide when a patch goes live. Book a free automation audit.
Questions people ask
What exactly did Google report about autonomous AI attacks?
Is my business at risk from this kind of attack?
When will these autonomous attacks become more common?
Should we stop using automation because of this report?
What is the first step a business owner should take?
Book a free automation audit
Thirty minutes. We look at one process you run every week and tell you exactly what an AI worker would take off your desk, and what it would not.

































