Saturday, Nov 21 | --:--
Back to home

Amazon: Attackers' AI Agents Found a MONAI Label Zero-Day; No Patched Release Yet

Amazon Threat Intelligence says AI agents run through the open-source Cairn framework by a suspected Chinese-nexus operator found a path-traversal zero-day in MONAI Label, an open-source medical-imaging AI application, and used it to get a root shell on a university server in September. The maintainers' critical advisory lists no patched version for 0.8.5 and earlier; the fix is merged but unreleased, and 0.8.5 was still the newest PyPI release when we checked on November 21. Attribution is Amazon's, at moderate confidence.

5 min read Seattle View as Markdown
A handmade editorial collage on torn ivory paper in cobalt and vermilion: a tall sheet of pale-blue paper film hangs upright from a single metal clip on a short wooden peg, a long vertical slit cut down its middle, with three soft cut-paper moths — plump furry bodies, broad fuzzy wings and feathery antennae — slipping along and out through the slit, one already drifting down, while an unused folded pale-blue paper patch still sits inside its open packet at the lower right; an official Amazon logo is glued on a torn-paper clipping whose top edge is tucked under the film's lower torn edge, with a soft shadow from the film falling on the clipping, the full logo visible; one vermilion scrap in the far corner.

The part of Amazon's report that matters to defenders is not that an AI agent found a bug. It is that the bug it found is still open. Amazon Threat Intelligence says AI agents run through Cairn, an open-source autonomous penetration-testing framework, by a suspected Chinese-nexus operator discovered a zero-day in MONAI Label, an open-source medical-imaging AI application, and used it to take full control of a university server. The maintainers merged a fix in October, but no release containing it has shipped.

What Amazon says happened#

In a post published on November 20, Amazon chief information security officer CJ Moses writes that on September 10 Amazon found a publicly hosted Cairn instance running without authentication. Because it was open, the operator's project configurations, prompts and the agents' output were visible. Amazon says the campaign targeted IP addresses tied to universities in East Asia outside mainland China over about 48 hours from September 8, and that the operator's goal for the agents was "obtain shell access."

Against the MONAI Label target, Amazon says, the agents identified the framework, mapped its API, found an upload endpoint with a filename parameter and tested it for path traversal. In about 11 minutes they went from mapping the API to writing arbitrary files on the server. In a later session they used the same flaw to plant a Python reverse shell that ran as root. Amazon says it assesses "with high confidence" that the flaw was a zero-day, because it found no prior fix, issue, advisory or database entry.

The agents did not get in everywhere. Amazon describes about 16 hours and 148 sequential actions against an open Redis server at one university, and a credential-spraying run of more than 30,000 combinations against a university GPU server, both unsuccessful. After the zero-day was found, Amazon says, the operator gave the agents a proxy into a second university network whose server ran the same software, and Amazon infers the operator had already used the flaw there.

Amazon attributes the activity to a Chinese-nexus actor with moderate confidence, citing Simplified Chinese prompts, China Standard Time timestamps in scripts and the choice of targets. It says the affected organisations were not AWS customers and that it worked with a government partner to notify them. The Register reported the same account from Moses's report on November 20.

The open exposure#

The MONAI Label advisory, GHSA-qc22-rmgc-vg94, rated critical and published on November 18, says the PUT /datastore/image endpoint uses a caller-supplied image ID without validation, so a value containing ../ can write an uploaded file anywhere the MONAI Label process can write. It says that can lead to remote code execution. Authentication is off by default, so in the default configuration anyone who can reach the server can exploit it; with authentication on, the ANNOTATOR role is enough.

The advisory lists versions 0.8.5 and earlier as affected, names no patched version, and says the fix, merged in pull request #1915, "will be included in the next MONAI Label release." When we checked PyPI on November 21, 0.8.5 was still the newest release. Until a new release ships, the advisory's workarounds are to turn authentication on and keep the ANNOTATOR role away from untrusted users, keep servers off untrusted networks, and run the server as an unprivileged user that can write only to its datastore. Because exploitation has been documented, VentureBeat also advises administrators to look for signs of an earlier compromise, since enabling authentication now will not undo one.

Claims and checks#

The agent behaviour, timings and attribution are Amazon's account of logs it read on the attacker's exposed server; we have no independent forensic record. The vulnerability itself is confirmed by the maintainers' own advisory, which says the issue was originally reported by an AWS security disclosure account.

Cairn has turned up on our file before: Reuters and NorForge described one operator using Cairn to breach more than 100 company websites, and Google's threat group reported the first fully automated exploit pipeline it had seen. Our read: the Amazon case moves the question from whether agents can find new bugs to how long a fix takes to reach users after they do. Here the code fix waited about a month for a release, and still does.

Limits#

  • The account of the campaign, the 11-minute figure and the Chinese-nexus attribution are Amazon's; Amazon rates its attribution at moderate confidence.
  • Amazon does not say how many MONAI Label servers are exposed online or whether other operators have used the flaw.
  • The advisory has no CVE identifier or CVSS score in its published record as we read it; we cite the GHSA id.
  • Amazon's post first appeared on November 20, before this edition's window; we include it because the exposure it describes was still unpatched in a release on November 21.

Sources#

Prior Coverage

Earlier Times of AI reporting on this thread.

Scroll to continue reading