Follow
Subscribe via Email!

Enter your email address to subscribe to this platform and receive notifications of new posts by email.

Amazon Discloses China AI Cyber Attack on University Server

Amazon reported that AI agents linked to China autonomously discovered a zero-day vulnerability in MONAI Label to breach an East Asian university server.
A person types on a laptop displaying code and a digital brain beside a Chinese flag.

Amazon disclosed on November 20 that AI agents linked to China found an unknown software flaw and took control of an East Asian university server. The breach unfolded during a September campaign that targeted multiple academic institutions outside mainland China. Threat intelligence teams at Amazon tracked the activity back to Cairn, an open-source tool coordinating agents for security tests. Rather than relying on rigid attack scripts, the agents pursued network access over dozens of hours on their own without needing step-by-step instructions from a person. The findings reveal how a China AI cyber attack moved from network scanning to root access with minimal human intervention [2].

China AI Cyber Attack Targets East Asian Universities

Amazon’s threat intelligence team found the attacker’s Cairn server on September 10. The machine was open to the web without passwords, leaving the operator’s prompt history in plain view for investigators during their review. Logs showed the campaign started on September 8 and ran for 48 hours [2].

Prompts on the server were written in Simplified Chinese, and scripts run by the tools matched China Standard Time, leading Amazon to assess a China connection with moderate confidence. The operator gave the tools a clear goal: gain shell access on target systems. Shell access gives an intruder the ability to run terminal commands on a remote host computer. Once active, the agents scanned university networks in East Asia outside mainland China to find open services, test perimeter defenses, and search for weak entry points across academic systems. Jessica Lyons reported in The Register that the incident shows how offensive testing tools can locate zero-day flaws during live operations conducted against real-world educational targets [1].

Security gatherings like Black Hat and DEF CON have tracked similar automated threats, where community leaders like Jeff Moss launched the DEF CON Franklin project to harden weak infrastructure. These defensive efforts mirror wider industry warnings that emerged after Google warned automated AI attacks could emerge to probe internet-facing systems across global computer networks. Security firms such as Palo Alto Networks have also examined how open-weight models influence offensive tactics [1]. Instead of halting when an attempt failed, the agents adjusted tactics and pursued alternate entry paths [2].

Adapting Without Human Direction

The automated campaign did not succeed on every system it encountered during testing. Against one university’s Redis server, the agents spent roughly 16 hours taking 148 sequential actions to break through security barriers and gain command access. They tested new passwords, changed settings, and tried loading unauthorized code. When security defenses blocked one route, the agents adjusted their approach and tested another method without waiting for human commands or stopping their automated routines. Could an agent keep guessing for a day? Server logs confirmed that the agents probed Redis configs and credentials for hours on end, though none of those attempts yielded access to the target host [2].

A separate operation targeted an academic GPU server for more than six straight hours. The agents gathered 80 real usernames from university web pages, published papers, and public campus records to run targeted password attempts against the server’s login portals. Research labs deploy fast hardware like Nvidia AI clusters, Qualcomm Dragonfly AI200 units showcased at SC26, and AMD 288 teraFLOPS MI430X systems, making academic GPU environments prime targets for computational hijacking [1]. Defensive controls on the university GPU machine stopped every intrusion attempt [2].

Editorial graphic illustrating a China AI cyber attack on server infrastructure.
Superpower Daily reported on Amazon’s investigation into autonomous agent activity targeting university infrastructure. (Credit: Superpower Daily)

Amazon CISO CJ Moses pointed to those failed attempts as evidence of how offensive agents now operate. “Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction,” Moses said. While earlier agent incidents involved research errors where OpenAI agents leaked user images online during testing, this campaign involved deliberate offensive operations conducted against live university targets. The agents did not pause; they simply chose another action [2].

Detecting the China AI Cyber Attack in Medical Code

The main breach hit MONAI Label, an open-source medical-imaging AI tool used in research. After spotting the tool, the agents mapped its open interfaces, found an upload feature that accepted custom filenames, and wrote files outside the storage folder in just 11 minutes [2].

The flaw discovered by the agents was path traversal. In MONAI Label, image and label upload functions assembled file destinations from user inputs without verifying whether the resulting paths remained safely inside the designated storage folder on the server. Because input checks were missing, the agents could write files to any location where the server process held write permissions. That gap gave attackers a way to overwrite app code or alter settings files. Amazon assessed with high confidence that this was a genuine zero-day flaw, as analysts found no prior advisory, public bug report, or database entry before the September campaign. The attackers used this weakness to establish their first foothold on the host machine [2].

Human operators still provided occasional assistance during the China AI cyber attack. In one session, the operator provided a network proxy that allowed the agents to enter a second campus network hosting the same medical-imaging app. Once inside, the agents located the upload interface and exploited the path traversal flaw a second time. They applied the technique at once, needing no further guidance to expand into the new environment [2].

Cybersecurity illustration published by The Register covering autonomous software breaches.
The Register covered security reporting on autonomous software tools targeting university infrastructure. (Credit: The Register)

Root Privilege Escalation on Host Systems

Writing files outside the storage folder was only the first phase of the operation. In a follow-up session, the agents used the file write flaw to place and run a custom Python script on the server. That script gave the agents interactive shell access under the root account, providing complete administrative control [2].

Root access gave the attackers control over the campus server and its connected storage drives. Because MONAI Label ran with root rights, the modified files executed with full authority on the host operating system. Jessica Lyons reported in The Register that offensive tools chained exploits together to take over the machine instead of stopping at simple reconnaissance or passive network scans [1]. The first 11-minute window measured file writes, but the complete takeover unfolded over several sessions as the agents confirmed their access and tested command execution on the host. Amazon reported that the agents systematically validated each step before running secondary payloads on the server [2].

VentureBeat reported on November 20 that administrators should inspect app files for unexpected changes, as tightening network controls after a breach does not undo existing server compromises. Once an intruder gains root access, simply restarting the service leaves modified files in place. Campus administrators must examine system logs and verify file integrity to ensure attackers did not leave persistent backdoors behind [2].

Security Guidelines for Vulnerable Servers

Amazon informed the open-source maintainers after uncovering the breach. Developers merged a fix into the main branch on October 19, and maintainers issued a security advisory on November 18 covering version 0.8.5 and earlier releases. The code update checks user-supplied destinations to ensure paths stay within permitted image and label folders, while also adding checks on incoming image identifiers. However, VentureBeat reported on November 20 that version 0.8.5 remained the newest release, meaning users must pull code from the source repository because a packaged release was not yet available. Until a new package ships, system administrators must manage the risk through local settings [2].

Authentication in MONAI Label is disabled by default. That setup allows any network client to reach the upload interface and trigger the path traversal bug. Turning on login controls limits general exposure, but accounts with the ANNOTATOR role can still trigger the flaw [2].

Maintainers outlined clear steps to protect medical AI systems. System teams should enforce login rules, restrict the ANNOTATOR role to verified lab members, isolate servers from open internet traffic, and remove root privileges from the host process. Restricting file write permissions on the storage folder also protects the host so that rogue uploads can’t overwrite core system binaries or alter settings files [2].

Sources
  1. ONLINE NEWS Lyons, J. (2026, November 20). China attackers deployed AI agents that autonomously found and exploited a 0-day to take over a university server. The Register. [Article Link]
  2. ONLINE NEWS Superpower Daily Editorial Desk. (2026, November 21). Amazon says attackers’ AI agents found an unknown flaw and took over a university server. Superpower Daily. [Article Link]
Leave a Comment

Related Posts
Total
0
Share