AI agents find a zero-day on their own and hijack university servers

According to Amazon, an attacker unleashed several AI agents on servers at East Asian universities. The agents found a previously unknown vulnerability in medical imaging software on their own, exploited it and gained root access. It took only about eleven minutes to get write access on the server.

Industry Nov 20, 2026 8 min read Source: Google News (EG)
Illustrative image, AI-generated

What's going on

On Friday, November 20, 2026, Amazon published a report from its threat intelligence team, which investigates cyberattacks. It was written by CJ Moses, Chief Information Security Officer (CISO) of Amazon Integrated Security. The report describes an attack seen as a new escalation: an attacker set several autonomous AI agents loose on university servers in East Asia. The agents found a previously unknown security flaw, a so-called zero-day, on their own and exploited it. That let them take over a university server with full administrator rights.

Zero-day means the software's developers didn't know about the flaw yet and so had zero days to fix it. The affected software was MONAI Label, an open-source application that uses AI to help annotate medical images, for example by outlining organs in CT scans. According to Amazon, the human behind the attack only set a broad goal: get access to the system. The agents planned and carried out the individual steps themselves.

How Amazon caught the attack

The attacks lasted about 48 hours and began on September 8, 2026. They targeted network addresses belonging to universities in East Asia, explicitly outside mainland China. Amazon found out because of a mistake by the attacker. On September 10, investigators discovered a publicly reachable server running an instance of Cairn with no authentication at all, meaning no password protection. There they could see the full history: which prompts the attacker had entered and which actions the agents had taken in response.

Cairn is an open-source framework for automating penetration tests. These are authorized attacks that security professionals normally use to find weaknesses in their own systems. You describe a goal to Cairn in plain language, and the tool splits the work across several AI agents. Moses says that makes it ideal for attackers. In this case, the attacker supplied IP addresses and asked the agents to "obtain shell access". That means a command line on someone else's machine that lets you run any command you want.

"Based on these inputs, the AI agents planned, executed, and adapted their approach as they learned, without the operator choosing each next step," Moses writes.

Eleven minutes to break in

According to Amazon, the successful attack went like this: the agents first identified which framework the application was built on. Then they mapped MONAI Label's API endpoints, the interfaces programs use to talk to the application. Along the way they spotted an upload function with a filename parameter. They tested that parameter for path traversal. With this kind of flaw, you sneak path elements like "../" into a filename and can drop files outside the intended folder. Think of a mail slot you can reach through to grab anything in the apartment. The test worked. According to Amazon, about 11 minutes passed from probing the APIs to writing arbitrary files on the server.

In a later session, the agents used the same flaw to plant code for a reverse shell. Here, the compromised machine itself opens a connection to the attacker, which many firewalls don't block. The code ended up in a file the Python interpreter loads automatically at startup. As soon as a function in the application was triggered, the connection was established, and the attacker had root access, the highest privileges on the system.

Amazon assesses "with high confidence" that the flaw was a zero-day. At the time of the attack there was no fix, no issue entry, no security advisory and no listing in vulnerability databases. The operator then gave the agents a proxy, a kind of tunnel, into the network of a second university running the same software. According to Recorded Future News, another break-in succeeded there. Amazon itself is more cautious: it believes the attacker exploited the flaw there too, but says this is an inference from the available evidence and not independently confirmed in every detail.

Hours of failure, without a break

Not every attempt succeeded, and the failed ones in particular show how persistent the agents are:

Target Duration Approach Result
University Redis server approx. 16 hours 148 consecutive actions failed
University GPU server over 6 hours 80 real usernames, more than 30,000 login combinations failed
MONAI Label approx. 11 minutes to write access path traversal, then reverse shell root access

The Register describes the Redis attempts in more detail. Redis is a fast database that keeps data in memory. The agents abused its replication feature to load foreign code, tried to break out of the scripting sandbox, wrote to configuration files and corrupted memory. After that they tried a Log4Shell injection against another service (Log4Shell is a well-known, severe vulnerability from 2021), a server-side request forgery against a Qdrant instance on the same machine, and several password attacks against PostgreSQL, SSH and other services. The agents had compiled the 80 usernames for the GPU server from the university's public web pages, research papers and other documents.

"Neither operation succeeded, but together they show the agents adapting their approach for almost an entire day without human guidance," Moses writes.

He warns that an agent that finds and exploits a zero-day in a few minutes gets its work done "faster than many security operations centers can manually triage an alert". The agents ran for 16 hours straight, with no breaks, no sleep and no working hours.

Who's behind it?

Amazon suspects a link to China with moderate confidence. The prompts were written in Simplified Chinese, the timestamps were set to China Standard Time, and the choice of targets fits. That proves neither who the attacker is nor that they work directly for the Chinese government. Amazon doesn't name the affected universities or their country. They were notified through a government agency.

This case isn't isolated. According to Recorded Future News, security firm Gambit Security had already documented in September that a financially motivated group was using Cairn against online retailers. According to The Register, a Chinese operator carried out at least 105 attacks in that campaign and compromised at least 27 companies to varying degrees, including a Fortune 500 hotel group and a major US airline. Google reported this week that Cairn had turned up in break-ins at companies in the US aviation industry. The FBI also warned about a similar tool called Artex. Eyal Sela of Gambit Security considers Cairn more significant than Artex because it's a general-purpose tool with no built-in cybersecurity structure:

"When a tool is more generic, it's less predictable, because it has less structure for how it solves a task."

Moses sums it up: this framework is being used for both espionage and criminal operations.

What MONAI Label operators should do now

Amazon notified the MONAI Label developers. On November 18, 2026, a security advisory rating the flaw as critical appeared on GitHub. All versions up to and including 0.8.5 are affected. The fix landed in the code with pull request no. 1915, but it's only set to ship in a later release. The advisory doesn't yet name an installable patched version. One important detail: authentication is turned off by default in MONAI Label. Until an update arrives, the project recommends three things: turn on authentication, block access from untrusted networks, and run the service with the fewest privileges possible. According to news.lavx.hu, affected institutions should also check their logs for unauthorized access, since servers like these may hold patient and research data. That can trigger reporting obligations, for example under the GDPR.

Our take

This is one of the best-documented cases so far of AI agents finding and exploiting a new vulnerability on their own in a real attack. The documentation only exists because the attacker ran their own tool unprotected on the internet. Still, it's worth keeping a level head. The flaw was fairly simple: an upload that doesn't validate filenames, in a niche application exposed to the internet without password protection. Against better-secured systems, the agents failed despite hours of attempts. What's really new is the speed and stamina, not some superhuman sophistication. For defenders, that means standard mistakes like exposed services without authentication will be found faster and more widely than before, and a manual response often comes too late. The attribution to China rests on circumstantial evidence that Amazon itself rates only with moderate confidence. And by Amazon's own account, the break-in at the second university isn't fully confirmed either.

Sources

#cybersecurity#ai agents#zero-day#amazon#china
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.