What 11 minutes actually measured

The number traveling with this story is 11 minutes, and it deserves a precise reading before it travels further.

Amazon's sentence is specific: the agents "identified the application's framework, mapped its API endpoints, and found an upload endpoint with a filename parameter. They tested that parameter for path traversal, and the parameter was vulnerable. In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server."

That is a real result. Finding an unvalidated filename parameter and proving you can write outside the storage folder is the hard discovery step, and an agent did it in minutes, without a human picking each move. But it is not "hacked a server in 11 minutes." The interval Amazon timed runs from API mapping to arbitrary file write; it does not report a total time to discover the flaw, and the root shell came in a later session, when the agents used the same flaw to plant a payload in a file Python loads automatically at startup, and then triggered a training endpoint to start a new Python process that ran it. How much later, Amazon does not say.

The same report also records what the agents could not do. Against a Redis server at another university, they worked for roughly 16 hours through 148 sequential actions, trying known techniques one after another, then moved on to other services on the same machine. Nothing worked. Against a GPU server, they spent more than 6 hours building a list of about 80 real usernames from the university's public pages and papers, then tested more than 30,000 credential combinations. They never got in. Persistence is part of the story. So is a full day of failure.

Read whole, the report supports a narrower claim than the headline version: given a target with a common flaw class, agents can find and prove the flaw very fast, and they will keep trying other doors for hours without supervision. That is enough to change defensive planning. It does not need the inflated version.

A default setting did the rest

The flaw itself is ordinary, which is the uncomfortable part. MONAI Label is a server research teams use to label medical scans and train imaging models. Its datastore upload endpoints take a caller-supplied image identifier and join it onto the storage folder without checking that the result stays inside that folder. Supply a path that climbs out of the folder and the uploaded file lands wherever the server process can write: application code, a settings file, an SSH key file. The advisory is blunt about where that leads: remote code execution.

What turned a common bug into an open door is a default. Authentication in MONAI Label is disabled by default. In the default setup, anyone who can reach the server over the network can call the upload endpoint, no account needed. Turning authentication on narrows the hole but does not close it: a user with the ANNOTATOR role, the ordinary working role for people labeling scans, can still exploit it.

This is the pattern worth keeping. The agents did not defeat a hardened system. They found a research tool that ships with its login off, on a university network, reachable. There are many servers like that, in imaging labs and everywhere else.

The fix is merged. No release carries it.

For anyone running MONAI Label, the practical fact is the patch state. Amazon notified the maintainers, who confirmed the flaw and merged a fix on October 19, 2026. The fix resolves the destination path inside the approved folders and rejects empty or dot-prefixed identifiers. It is, by the advisory's own wording, to be "included in the next MONAI Label release."

That release has not happened. As of November 21, 2026, the advisory lists no patched version, and the newest installable release on PyPI is still 0.8.5, uploaded in November 2024, which is inside the affected range. A merged fix in a source tree is not a package a hospital research IT team can deploy.

Until a release ships, the advisory's interim mitigations are the defense for affected deployments: turn authentication on and keep the ANNOTATOR role to trusted users, keep the server off untrusted networks, and run it as an unprivileged user that can write only to its datastore. Anyone who ran an exposed MONAI Label server should also review it for unexpected file changes and unauthorized activity.

Two limits on the story should stay attached to it. Amazon assesses the Chinese-nexus attribution with moderate confidence, on language, time-zone and targeting evidence; it names no group and establishes no government direction. And nothing in the report shows patient data taken or a clinical system touched. The targets were university research servers. That distinction matters, because "AI agents breached medical software" will be quoted in rooms where the distinction gets dropped.

Sources

  1. [1] Amazon Threat Intelligence, "Chinese-nexus actor's Cairn AI agents found a zero-day on their own", AWS Security Blog, November 20, 2026Read source
  2. [2] Project-MONAI / MONAILabel, GitHub Security Advisory GHSA-qc22-rmgc-vg94, "Path traversal in datastore upload endpoints allows arbitrary file write", published November 18, 2026Read source
  3. [3] The Hacker News, "Attackers' AI Agents Found a Zero-Day in Medical-Imaging Software and Gained Root Access", November 21, 2026Read source
  4. [4] VentureBeat, "Amazon says attackers' AI agents found and exploited a zero-day in 11 minutes, but no patched release is available", November 20, 2026Read source
  5. [5] The Register, "China attackers deployed AI agents that autonomously found and exploited a 0-day to take over a university server", November 20, 2026Read source
  6. [6] CVE record CVE-2026-133968Read source