Hackers' AI Found a Hole Nobody Knew About and Broke In Within 11 Minutes. What on Your Server Is Open to the Internet?
If you run a website, an admin page, or any tool on a server that can be reached from the internet, this is a check worth doing this week. Amazon Threat Intelligence reported on 20 November 2026 that attackers used auton
If you run a website, an admin page, or any tool on a server that can be reached from the internet, this is a check worth doing this week. Amazon Threat Intelligence reported on 20 November 2026 that attackers used autonomous AI agents to find a hole nobody had reported yet in an open-source medical-imaging web app. BleepingComputer, writing about the same research, says the agents broke in within 11 minutes: they went from mapping the app's public interface to using that unknown hole. They later used the same hole to take full control of a university server, and then against a second university.
Amazon's team saw the whole campaign because the attackers left their own dashboard open to the internet with no password.
What happened
On 10 September 2026, Amazon Threat Intelligence found a public copy of Cairn, an open-source tool that can coordinate several AI agents to reach a goal written in plain language. The copy was reachable from the internet and had no login, so anyone could see the operators' prompts and the agents' work history.
The attackers had given the agents lists of internet addresses and a simple instruction: "obtain shell access," meaning find a way in and take control of the machine. According to CJ Moses, CISO of Amazon Integrated Security, the agents then planned, carried out, and adjusted their own steps. The operator did not choose each next move. Amazon says the campaign ran for about 48 hours from 8 September 2026 and targeted universities in East Asia. The universities were not named.
Not every attempt worked. One effort ran 148 actions over 16 hours and failed, including a long try against a database service. Moses said those agents "ran continuously for 16 hours without breaks, sleep, or working-hour constraints." In a separate attempt, the agents spent more than six hours trying more than 30,000 password combinations for about 80 real usernames, including against a remote login, an admin dashboard and a VPN, and never found a working login.
One attempt did work. The agents found a previously unknown flaw in an open-source medical-imaging AI web app. BleepingComputer identifies that app as MONAI Label, a tool researchers use to label medical images. Amazon assesses with high confidence that the flaw was unknown at the time: there was no public fix or advisory. In about 11 minutes, the agents went from mapping the app's public interface to using that hole. In a later session, they used the same hole to take full control of a university server, then against a second university running the same software.
Amazon contacted the maintainers, and they released a fix. BleepingComputer says that fix came out on 18 November 2026. AWS also worked with a government partner to notify the affected organizations, which were not AWS customers. Amazon assesses with moderate confidence that the actor was China-linked, based on the language of the prompts, the time zone on the scripts, and the choice of targets. This article follows Amazon's own wording and does not treat that as proven.
Who is affected
The victims in this report are universities, and they have not been identified in public. The wider lesson is not limited to universities or to medical-imaging software. Anyone who runs a website, an admin panel, an API, a dashboard, or a self-hosted open-source tool that can be reached from the internet can face the same kind of attention. If you run MONAI Label, update it. If you run other tools on a server, assume a stranger can find them if they are left open.
Why it matters for a small business
You may not think of your login page, your admin panel, or a leftover test app as "open to the internet." If they can be reached without being on a private network, they are. In one case, the agents found a hole nobody knew about and moved in minutes. In others, they spent hours trying passwords and still failed, which is a reminder that rate limits and lockouts matter. The attackers themselves were caught because they left a dashboard open with no password.
Moses warned that an agent that finds and uses an unknown hole in minutes can finish its work faster than many security teams can even review an alert. A small business often has the same pattern: a website, a store or WordPress admin page, a hosting control panel, or an old tool nobody uses anymore, all sitting on a server.
What to check today
1. Check which apps, admin panels and APIs are reachable from the internet
Ask your developer or host for a simple list: what on your server can be opened from the public internet? Include admin pages, customer logins, APIs, dashboards, staging copies, and any self-hosted tools. If you do not recognise an item, or nobody uses it, treat it as a problem until you know why it is there. A free website security scanner can show what an outsider can see from the internet, without logging in. It is not a full review of your server, and it will not by itself prove you are safe.
2. Take down what you do not need, and update the rest
Turn off or remove tools, admin pages and test copies you do not need. For what you still use, apply the latest updates, especially for self-hosted open-source apps. The maintainers of the medical-imaging app in this case released a fix on 18 November. We have also covered cases where a service was left reachable and then abused, including SQL Server logins left open on the internet and web-server flaws that can let attackers in without a password. If it does not need to be public, it should not be public.
3. Turn on login rate limits and lockouts
The agents tried more than 30,000 password combinations for 80 real usernames over six hours and failed. That is the kind of guessing that rate limits and account lockouts are meant to stop. If your site, admin page, email, or hosting panel can be tried again and again with no pause, ask your developer to turn on limits and lockouts. Our guide to brute-force login protection walks through what to ask for. A related case is a password-reset form that attackers guessed their way into.
4. Never leave a dashboard open without a password
Amazon found this campaign because the attackers left their own Cairn dashboard reachable from the internet with no login. The same mistake is easy to make with a hosting panel, a store admin page, a database tool, a test site, or a monitoring dashboard. Every dashboard should require a strong login. If a page does not need to be on the public internet, keep it off the public internet. AI agents already look for things left in the open, as we have written about how they hunt exposed API keys and about Gemini agents hitting a website tens of thousands of times.
FAQ
Did they really break in within 11 minutes?
Amazon and BleepingComputer both describe an 11-minute window in which the agents broke in within 11 minutes, going from mapping the app's public interface to using a hole nobody had reported yet. Full control of the university server came in a later session, when they used that same hole again.
Did all of the AI attacks work?
No. One attempt ran 148 actions over 16 hours and failed. Another tried more than 30,000 password combinations for about 80 real usernames over six hours and also failed. The successful case was the unknown flaw in the medical-imaging app.
How did Amazon see the attacks?
The attackers left their Cairn dashboard open to the internet with no password. Amazon could read the prompts and the work history. That is also the simplest warning in the report for anyone who runs a dashboard of their own.
Is there a fix, and is there a CVE?
The maintainers released a fix. BleepingComputer dates that release to 18 November 2026. Amazon says there was no public advisory at the time the agents found the hole. This article does not list a CVE or advisory number. If you run MONAI Label or any other self-hosted tool, update it and check what else on the server is reachable from the internet.
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.