Osegi
Security·2 min read

MONAI Label zero-day: AI agents hacked a university server

Amazon says an attacker gave AI agents a goal and they found a zero-day in MONAI Label on their own, then took over a university server in East Asia.

A server room with a glowing red warning screen representing an AI-driven attack on a university server.
Image: البوابة التقنية

An attacker pointed a swarm of AI agents at university servers in East Asia and told them, in plain language, to get command-line access. The agents did the rest: they found a previously unknown flaw in the medical imaging tool MONAI Label and used it to take control of a machine. That is what Amazon's threat intelligence team says it found.

According to a report by Amazon's chief information security officer, CJ Moses, the attacks ran for about 48 hours starting 8 September 2026. They targeted network addresses tied to universities in East Asia outside mainland China. Amazon's team spotted the activity on 10 September, when it found a server running a platform called Cairn with no authentication, which exposed the attacker's full log of instructions and actions.

What is Cairn, and how did the attack work?

Cairn is an open-source tool for automating penetration tests. It coordinates several AI agents toward a goal written in natural language. Here, the attacker supplied the target addresses and asked for command-line access. The agents planned steps, tried them and adjusted after each result.

MONAI Label is open-source software that uses AI to help label medical images. The agents probed its APIs and found a flaw in how file paths are checked in the upload function. About 11 minutes after they began scanning the APIs, they could write files in places they shouldn't. In a later session, they planted code in a file that the Python interpreter loads automatically at startup. Running one of the app's functions then opened a reverse connection that gave the attacker root privileges.

Not everything worked. One run lasted about 16 hours and 148 steps against a Redis database server, without success. Another, over six hours against a GPU server, gathered about 80 usernames and tried more than 30,000 credential sets, and still failed to get in.

Is there a fix for MONAI Label?

Amazon told the project, which published a security advisory on GitHub on 18 November 2026 rating the flaw critical. Versions up to 0.8.5 are affected. A code fix exists in pull request 1915, but it will ship in a later release, and no patched version is available to install yet.

Authentication is off by default in the app. Until a fix arrives, the advisory recommends turning it on, blocking untrusted networks from reaching the server and running the service with limited privileges.

Amazon rates with medium confidence that the operator is linked to China, citing Simplified Chinese instructions, China-based timing and the choice of targets. It says this is not proof of identity or of government ties. It did not name the universities and says it alerted a government agency.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:amazonai agentszero daymonai labelcairncybersecurity

More from this day

See all →

More in: Security

See all →