AI Agents Found Their Own Zero-Day — and Breached a Medical Imaging Server

Autonomous AI agents have crossed a line security researchers have been warning about for years: they found a zero-day vulnerability nobody knew existed, exploited it, and took root access on a real server — with almost no human guidance.

Amazon Threat Intelligence disclosed the incident this week. Suspected China-linked hackers used AI agents running on Cairn, an open-source framework that runs AI agents as automated penetration testers, to find a previously unknown flaw in MONAI Label, an open-source medical imaging server, and gain root access on a university server. The Hacker News reported the disclosure on November 21, 2026.

This is not a lab demo. It is a documented intrusion, reconstructed from the attacker’s own exposed control panel — and it changes the math on how fast an unknown vulnerability can go from undiscovered to exploited.

What Happened

Over about 48 hours in September, the AI agents mapped the medical imaging application’s interface, probed it, and reached the point of writing files to the server in roughly 11 minutes, according to Amazon’s incident report. The operator set the goal — get shell access — and chose the targets. The agents selected the methods themselves.

Amazon reconstructed the intrusion after finding the operator’s own Cairn control panel sitting on the internet without a password. The exposed panel revealed the agents’ task settings, prompts, and run history — an operational security failure by the attackers that handed investigators the full picture.

The target was MONAI Label, a server that research teams use to label medical scans and train imaging models. It is a separate program from the broader MONAI framework.

The Flaw: CVE-2026-133968

The weakness is a path-traversal flaw in the parts of the server that store uploaded images and labels. A file identifier supplied by the caller is appended to the storage folder and written to disk without checking that the result stays within that folder.

The flaw, now listed as CVE-2026-133968 and rated 9.8 out of 10 — a critical score — affects MONAI Label 0.8.5 and earlier. The default setup runs with no login enabled, so any client that can reach the server over the network can trigger the write. The server writes with the privileges of the account it runs under, so an attacker cannot write anywhere on the machine — but a writable code or settings location can open a path to running code.

The root shell came later, in a separate session. The agents used the same flaw to plant a file that Python runs automatically at startup, then triggered a training feature that started a new Python process. That process loaded the planted file and connected back to the attacker as root — on one university server. Whether another MONAI Label system would give the same result depends on how it is run and what its service account can do.

In a later session, the operator handed the agents a tunnel into a second university network running the same software. Amazon believes that network was broken into the same way but has not confirmed the second compromise.

Who Was Behind It

Investigators tie the campaign to a China-linked actor with moderate confidence, pointing to Simplified Chinese prompts, China Standard Time settings, and targets likely of interest to China. They did not connect the work to a named group or establish that a government was behind it.

Amazon’s account is based on its own reconstruction and has not been independently confirmed. It does not show that patient scans or records were taken, that clinical systems were touched, or that the attacker reached past the university targets described. The campaign also included failed operations, and the root result on one server should not be taken as the outcome for every MONAI Label deployment.

The Patch That Isn’t There Yet

Here is the part that should worry anyone running this software: there is still no fixed version to install.

The maintainers confirmed the flaw and merged the source fix on October 19, which forces uploads to resolve inside the approved folders and rejects empty names or names that begin with a dot. But their security advisory, published earlier this week, still lists no patched version and says the correction will arrive in a future release. The newest release listing remains 0.8.5, which dates to 2024. A merged fix in source code is not the same as a package you can install.

Until a patched release ships, the practical defenses are to cut off exposure and limit what a compromised server could reach: block untrusted network access to MONAI Label servers with firewall and routing rules, turn on authentication (which the server leaves off by default), run the server under a limited account rather than root, and restrict the folders it can write to. Do not assume a source build is safe — confirm it includes the October fix before treating it as patched. And if a server was reachable, check for unexpected Python startup files and unusual training jobs, the path the reported attack used.

BADBOX 3.0: The Other Malware Story This Week

The AI-agent intrusion was not the week’s only major malware news. HUMAN Security announced on November 18, 2026 that its Satori Threat Intelligence and Research Team had disrupted BADBOX 3.0, the largest ad fraud operation by bid volume ever publicly disclosed.

At its height, BADBOX 3.0 generated roughly 40 billion fraudulent advertising bid requests a day — more than three times the volume of the previous record holder, VASTFLUX — and accounted for more than 4% of all connected TV traffic observed by the Human Defense Platform. The operation ran primarily on low-cost, uncertified Android TV boxes and streaming devices, infected with a backdoor packaged as a system application that fetched fraud modules from public code repositories. The modules rendered video ads the device owner never saw, borrowing the identities of legitimate apps — and the same infected boxes quietly rented out their owners’ home internet connections as residential proxies, without consent.

Satori researchers linked the BADBOX 3.0 module catalog to TXCZ, a Chinese electronics manufacturer, through server infrastructure, a factory-installed launcher, and code that checked for TXCZ-specific device settings. A separate Google-led takedown of the XCPE botnet also hit several domains that controlled parts of BADBOX 3.0; Google said XCPE had enrolled more than 1.3 million set-top boxes.

The wider threat picture stayed busy as well: CISA added seven vulnerabilities to its Known Exploited Vulnerabilities catalog this week, and ransomware leak sites posted 379 new victims over the last seven days, according to ThreatPodium’s November 21, 2026 threat intelligence report.

The Bottom Line

Two stories, one lesson. The BADBOX 3.0 takedown shows malware operations now run at industrial scale — 40 billion fake ad requests a day, generated by TV boxes sitting in people’s living rooms. The MONAI Label intrusion shows the other end of the spectrum: AI agents that can find a vulnerability nobody knew about, figure out how to exploit it, and take root — in minutes, with a human operator who only had to point them at the target and wait.

The window between “unknown flaw” and “exploited in the wild” used to be measured in weeks or months, and it depended on a skilled human finding the bug first. That window is now measured in minutes, and the finder may be software. Patching faster is no longer enough on its own — the software with no patch available, the server with authentication off by default, and the device that was compromised before it was plugged in are the exposures that matter now. Audit what is reachable on your network, turn on the authentication that ships off, and assume anything internet-facing will be probed by something that never sleeps.

Sources: The Hacker News, “Attackers’ AI Agents Found a Zero-Day in Medical-Imaging Software and Gained Root Access” (November 21, 2026); Amazon Threat Intelligence incident report, via The Hacker News; HUMAN Security, “HUMAN Security Disrupts BADBOX 3.0 Ad Fraud Operation” (November 18, 2026); HUMAN Security Satori Threat Intelligence report on BADBOX 3.0; BleepingComputer, “BADBOX botnet returns with record-breaking 40 billion daily ad requests”; ThreatPodium, “Threat Intelligence Report — November 21, 2026.”

Edited by Muse @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous
Never miss a customer call again — MAJ.COM’s AI receptionist answers every call, texts customers back instantly, and books the appointment for you. See how it works at https://MAJ.COM


Discover more from QUE.com

Subscribe to get the latest posts sent to your email.

Leave a Reply

Connect with

Discover more from QUE.com

Subscribe now to keep reading and get access to the full archive.

Continue reading