
Photo: NOIRLab/NSF/AURA/T. Slovinský, 2022, via Wikimedia Commons (CC BY 4.0)
For years, security researchers warned that AI would eventually run hacking campaigns largely on its own. In the space of five days in November 2026, three of the biggest names in cyber defense said that moment has arrived. On Nov. 13 the FBI issued a warning about ARTEX, a free, open-source tool that strings AI agents together to break into networks. The same day, Amazon described AI agents it caught stealing AI access keys to pay for their own work. Then on Nov. 16, Google reported what it called the first case of attackers using AI to find and exploit software flaws "without a human in the loop."
The short answer: AI hacking agents are programs that take a goal written in plain English, such as "get into this company’s cloud account," and then plan and carry out the steps themselves by calling a large language model like the ones behind ChatGPT, Claude or DeepSeek. The FBI says attackers have used ARTEX since at least August 2026 and that some intrusions went from first contact to administrator access in under three hours. Google found a separate system, which it calls exploit_agent, that compared old and new versions of popular web software, worked out what a security patch had fixed, wrote an exploit and fired it at websites with no person approving each step. The attacks mostly use familiar weaknesses (unpatched software, exposed admin panels, default passwords, leaked keys), so the defenses are familiar too. What has changed is the speed.
Below: what each report found, how an AI-run intrusion unfolds step by step, the numbers that matter, a reality check on the scarier headlines, and what website owners, AI builders and everyday users should do now.
Three warnings in five days
The FBI, Nov. 13. The bureau’s FLASH alert (number FLASH-20261113-001) describes ARTEX as "an open-source, large language model (LLM)-driven agentic platform that threat actors have weaponized to automate multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration." According to the FBI, a developer using the name "Autumn-27" released it on GitHub in July 2026 for authorized security testing, and it collected an estimated 1,500 to 2,000 GitHub stars in 10 weeks. After reports of abuse, the developer made the repository unavailable on Oct. 8, but the FBI says the tool lives on through forks, archived copies and a commercial version called ARTEX PRO. Targets have included financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications and legal organizations.
Amazon, Nov. 13. Amazon’s threat intelligence team runs MadPot, a worldwide network of decoy servers. From Oct. 3 to Nov. 5, some of those decoys posed as AI model services, and AI agents that Amazon believes, with moderate confidence, were running on ARTEX sent them requests carrying their own campaign notes. Those notes described a loop in which agents searched the internet for exposed AI tools, pulled out the upstream provider keys stored inside, tested them, and fed working keys back into the system to pay for more searching. A second campaign used real-time deepfake video and forged ID photos to try to beat the identity checks at two cryptocurrency exchanges. Amazon stresses that it saw the agents’ records, not the attacks themselves, and could not confirm the outcomes they reported.
Google, Nov. 16. Google Threat Intelligence Group (GTIG) and its Mandiant incident response unit published their quarterly AI threat tracker under the headline "Agentic Threat Activity Crosses a Significant Threshold." The centerpiece is exploit_agent, a multi-agent framework found on an attacker’s exposed staging server. Google says it connects to several commercial and locally run language models, watches popular web applications for security updates, compares the code before and after each patch, drafts an exploit for the flaw the patch reveals and runs it automatically. The server also held stolen victim data and scripts for scanning internal networks, guessing passwords, planting webshells (small backdoor programs on a web server) and deleting rival attackers’ webshells. Comments in the code were written in Russian, and victim records pointed to French media, entertainment and education websites. Google did not name a group or a country behind it.
How an AI-run intrusion unfolds, step by step
None of the three reports describes a single super-hacker AI. They describe a team of narrower agents, each handed one job, coordinated by software and powered by rented access to a language model. Put together from the FBI, Amazon, Google and Censys findings, a typical run looks like this.
Step 1: The operator writes the goal in plain English
The FBI says ARTEX "translates high-level tactical goals described in natural language into actionable machine commands" with "minimal human interaction beyond initial objective setting." Amazon’s captured records show operators writing standing instructions, such as which countries to work through in order, which targets to skip and how to grade each stolen credential, and then leaving the agents to it. ARTEX has no AI model of its own. It plugs into outside models, including Anthropic’s Claude and OpenAI’s GPT as well as DeepSeek, GLM, Grok and Kimi. In the attacker setups the FBI analyzed, DeepSeek’s low-cost Flash models were the most common choice because of their price and what the bureau calls a "lack of observable guardrails when framed for penetration testing."
Step 2: Agents map what is exposed
Scouting agents scan for internet-facing systems, test what stolen cloud credentials can reach with sub-second permission checks, and fire hundreds of requests at websites looking for hidden pages. Ways in listed by the FBI include SQL injection, web APIs with no authentication, weak admin passwords and administrator passwords hardcoded into a site’s JavaScript. The FBI notes that ARTEX is especially good at finding the systems companies forget about: partner and supplier portals, mobile app back ends, employee apps, sales tools, legacy services and apps built by individual business units. Amazon’s records show the key-stealing agents also combing GitHub, npm and PyPI packages, container images, notebooks, paste sites and Telegram channels for leaked secrets.
Step 3: A security patch becomes a roadmap
This is the step Google calls new. When a vendor ships a fix, the code change itself hints at where the flaw was. Exploit_agent automated what human researchers call patch diffing: compare the patched and unpatched code, infer the bug, write a working attack and try it on sites that have not updated yet. One screenshot in Google’s report shows the framework inspecting the access controls of Spectra, a popular WordPress page-builder plugin also known as Ultimate Addons for Gutenberg, version 2.19.25. That is the kind of add-on found on thousands of small business sites.
Step 4: Agents climb to administrator access
With a foothold, other agents try to escalate. The FBI’s table of observed ARTEX activity reads like a penetration tester’s checklist run on fast-forward: launching cloud servers that carry powerful roles, harvesting Windows service tickets, dumping database passwords through SQL injection, creating accounts flagged as administrators and forging admin tokens with keys that had never been rotated. In one intrusion the agents enumerated more than 110,000 Active Directory objects. In the fastest cases, the bureau says, ARTEX intrusions went from initial targeting to privileged access in under three hours.
Step 5: Data and keys go out the door
The last agents collect what the operator wanted: customer records, cloud keys, or access to paid AI models that can be resold or used to run the next campaign. The FBI’s two headline cases show the range. In attacks on South Korean financial institutions, personal information on about 67,000 people was taken. In a supply chain attack on ConnectWise Cloud Backup customers, the attackers attempted sign-ins to more than 2,250 downstream customer tenants.
ARTEX significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours. (FBI FLASH-20261113-001, Nov. 13, 2026)
The numbers that matter
Under 3 hours. The FBI’s figure for the fastest ARTEX intrusions, from first targeting to privileged access.
1,465 consoles. Internet scanning firm Censys counted 1,465 hosts with an exposed ARTEX control panel on Nov. 3 and 4. About 650 sat in two blocks of identically configured servers, which leaves roughly 815 distinct-looking setups. The FBI, citing researchers, puts the count of exposed consoles above 1,600 and says the true number is likely much higher because many run behind firewalls.
155 hosts. Of the ARTEX servers Censys found, 155 also ran other attack tools, most often a second AI hacking framework called CyberStrikeAI.
22 zero-days in one month. Google counted 22 exploited zero-days, flaws attackers used before a fix existed, in August 2026, up from an average of 11 a month earlier in 2026 and eight a month in 2025. Overall vulnerability disclosures and exploits nearly doubled between January and August.
Four days. After two flaws in PaperCut print management servers were disclosed in late August, researchers at Huntress used AI to produce working exploits within a day, and GreyNoise later reported an attacker who likely used AI to build exploits and hit hundreds of organizations four days after disclosure.
$100,000 in a day. Mandiant handled "LLMjacking" cases this year in which attackers used stolen cloud or AI credentials to run up more than $100,000 in model charges in a single day, and in the worst cases about $1 million before anyone shut them off. Google found underground resellers selling stolen model access for 10 cents to $25 per million tokens.
What is new, and what is not

Photo: Tony Webster, 2016, via Wikimedia Commons (CC BY 2.0)
Machines hacking machines is not a new idea. On Aug. 4, 2016, DARPA staged the Cyber Grand Challenge at the DEF CON conference in Las Vegas, where seven fully automated systems found, patched and exploited software flaws against each other with no human help. A system called Mayhem, from the Pittsburgh company ForAllSecure, won the $2 million top prize. Those systems were built for a contest and ran on high-performance computers inside a closed arena.
What changed in 2026 is who can do it, and against what. ARTEX is free to download, uses models anyone can rent, and points at the real internet. The FBI says it lets attackers "with minimal technical expertise" run complex campaigns against several targets at once. Google lists a whole family of similar open-source frameworks seen in real intrusions or experiments, including CAIRN, CyberStrikeAI, PentAGI and HexStrike AI, the last of which uses the Model Context Protocol to hand an AI agent more than 150 hacking tools.
The underlying weaknesses are old. Google’s analysts make the point directly: "the mechanics of the intrusions remain the same, such as using compromised credentials and targeting unpatched systems." The agents are fast at finding simple flaws and stitching steps together. They are not, on the evidence in these reports, inventing new classes of attack.
The agents are fast at finding simple flaws and stitching steps together. They are not, on the evidence in these reports, inventing new classes of attack.
A reality check on the headlines
"The first fully AI-run cyberattack." Mostly accurate, with a caveat. Google’s exact claim is that it saw attackers "fully automate vulnerability discovery and exploitation without a human in the loop for the first time." A person still set up the system, chose the kind of software to watch and supplied the AI keys. The no-human part covers the run from patch to exploit to attack.
"AI superintelligence is hacking us." No. The FBI’s alert calls ARTEX’s components "super intelligence (SI) agents," but the same document says the tool "does not include its own SI model" and simply calls commercial or open models. The capability comes from ordinary language models wired to ordinary hacking tools.
"Defenders cannot keep up." Too strong. Google notes that today’s frameworks "trade stealth for velocity": they generate heavy, noisy traffic and leave code traces that make detection more likely. The FBI likewise says ARTEX is "forensically noisy." The problem is that by the time an alert fires, a three-hour intrusion may already be finished.
"It gets worse from here." Probably. BleepingComputer reports that Google’s researchers expect unsophisticated attacks needing "little to no human interaction" within about three months and "sophisticated, fully autonomous" agentic attacks within about six. Google’s report says it expects less human involvement over time along with more stealth.
"Nobody can stop the tool now." Partly true. Taking down the GitHub repository did not remove ARTEX from circulation, according to the FBI. But Censys found many operators careless enough to leave consoles and file directories open to the internet, including 32 exposed directories holding attack kits, which gives defenders indicators to block.
What to do now
If you run a website, including a WordPress site
Turn on automatic updates for plugins, themes and the core software, or check for updates at least weekly. Google’s case shows the gap between a patch shipping and an AI-written exploit arriving can now be days or less. Remove plugins you no longer use, keep admin logins behind strong unique passwords and two-factor sign-in, and do not leave staging copies or admin panels open to the public internet. If your host offers a web application firewall, switch it on.
If you run AI tools, model gateways or agent apps
Amazon’s records show attackers specifically hunting LLM gateways and AI app builders, including LiteLLM, One API, Dify, Langflow, RAGFlow, Open WebUI and Ollama, because one weak dashboard can expose many provider keys. Change default admin passwords (the default LiteLLM master key was in the attackers’ playbook), patch known flaws such as CVE-2025-3248 in Langflow, keep management pages off the public internet, never ship keys in client-side code or public repositories, set spending limits and billing alerts with your model providers, and rotate any key that has ever been exposed.
If you manage a company network
The FBI’s own checklist starts with an honest inventory of everything that faces the internet, especially the forgotten portals and side projects ARTEX favors. It recommends restricting cloud admin, backup and deployment interfaces to trusted networks, applying least privilege to app permissions, watching for new credentials added to existing apps, rate-limiting sensitive business functions, hiding detailed error messages from outsiders and scanning code and build pipelines for embedded secrets. Signs of an AI agent at work include compressed bursts of activity, large numbers of failed attempts with small incremental changes, and signs of several sub-agents overwriting each other or sending duplicate requests. The bureau asks victims to report incidents to a local FBI field office or to IC3.gov.
If you are an everyday user
You are unlikely to be targeted by ARTEX directly, but your data sits in the portals and back ends these agents probe. Use a password manager and unique passwords so one breach does not open other accounts, turn on two-factor or passkey sign-in wherever offered, keep your phone, computer and router updated, and be wary of breach-related emails that ask you to "verify" an account.
What we still do not know
How many victims there are. The FBI lists campaigns and sectors but gives no total. Amazon could not verify what the agents it observed actually achieved, and Google did not say how many French sites exploit_agent hit.
Who is behind most of it. The FBI and Google describe a mix of financially motivated criminals and suspected state-backed groups, and attribution in several cases rests on clues such as language settings in code and prompts.
How well model safeguards work. ARTEX can use major commercial models, yet the FBI found attackers mostly chose cheaper models with weaker guardrails. Google, for one, says it monitors for this kind of misuse and disables offending accounts, but none of the reports measures how much gets through.
The bottom line
AI has not handed attackers a new superweapon. It has handed them a tireless team that can scan, test, write exploits and move through a network faster than most organizations can respond, using weaknesses that have existed for decades. The fixes are unglamorous and well known: patch quickly, close exposed admin panels, retire default passwords, lock down keys, and watch for activity that moves at machine speed. The difference in 2026 is that the window for doing them has shrunk from weeks to hours.
Further reading

This Is How They Tell Me the World Ends
A former New York Times security reporter traces the market for zero-day flaws, the bugs attackers prize most, and why that market keeps growing.

The Art of Invisibility
Once the most wanted hacker in the United States, Kevin Mitnick lays out practical steps for protecting accounts, devices and personal data.

Norton 360 Deluxe (5 Devices)
Antivirus, a password manager, cloud backup and a VPN for up to five computers and phones in one yearly subscription.

Amazon eero 6+ Mesh Wi-Fi Router
A Wi-Fi 6 router that installs its own firmware updates automatically, one less unpatched device on a home network.

Samsung T7 Portable SSD (1TB)
A pocket-size USB-C drive rated up to 1,050 MB/s, handy for keeping an offline backup that an intruder on your network cannot reach.
