Skip to main content
Emerging ThreatsMalware & Ransomware

AI Agents Exploit Zero-Day in Medical Imaging Software for Root Access

Medical imaging lab with workstations, equipment, and researcher.

CVE-2026-133968, a newly assigned vulnerability rated 9.8 out of 10, permits an unauthenticated attacker to write files outside MONAI Label’s storage folder and — in at least one confirmed case — helped an intrusion reach root on a university server, Amazon Threat Intelligence reported this week.

CVE-2026-133968 and the scope of the flaw

The weakness is a path-traversal bug in MONAI Label versions 0.8.5 and earlier when run in the default configuration, which leaves authentication disabled. The server appends a file identifier supplied by a caller to the storage folder and writes it to disk without checking whether the resulting path remains inside that folder. That allows any client that can reach the service over the network to cause files to be written outside the intended storage area. The server writes using the account privileges of the process that runs it; therefore the impact of a write depends on what that service account can do and where it can write.

How Cairn and autonomous agents discovered the bug

Amazon reconstructed the intrusion after finding an exposed Cairn control panel on the internet with no password. Cairn is an open-source framework that runs AI agents as automated penetration testers; the exposed panel contained the agents’ task settings, prompts, and run history. Over roughly 48 hours in September, the agents mapped MONAI Label’s interface and reached the point of writing files to the server in about 11 minutes, according to Amazon’s incident report.

The operator set a high-level goal — getting shell access — and chose the targets; the agents selected methods autonomously and required little step‑by‑step direction. Investigators tie the campaign to a China-linked actor with moderate confidence, citing Simplified Chinese prompts, China Standard Time settings, and targets they judged likely to interest China. Amazon did not attribute the activity to a named group or to a government.

The exploitation chain that led to root on a university server

The initial automated session produced file-write capability within minutes but did not immediately yield root. In a separate session, the agents again used the path-traversal flaw to plant a file that Python runs automatically at startup. They then triggered a training feature in MONAI Label that started a new Python process; that process loaded the planted file and connected back to the attacker as root. Amazon confirmed root access on one university server. Whether other MONAI Label deployments would produce the same result depends on how the service is run and what its account can access.

Amazon also described a later session in which the operator provided the agents a tunnel into a second university network running the same software; Amazon believes that network was compromised in the same way but has not confirmed the second compromise. The company’s account is based on its reconstruction and has not been independently confirmed. Amazon’s report 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 included failed operations, and the root result on one server should not be taken as inevitable for every deployment.

What MONAI Label maintainers changed (and what hasn't shipped)

Maintainers merged a source-code fix on October 19 that forces uploads to resolve inside approved folders and rejects empty names or names that begin with a dot. However, the project’s security advisory — published earlier this week — still lists no patched release; the newest release listed remains 0.8.5, dated to 2024. A merged fix in source control is not the same as a packaged release that users can install, and the advisory notes the correction will arrive in a future release. The CVE record for the flaw (CVE-2026-133968) was published this week and rates the bug 9.8/10; the advisory’s lack of a CVE on the project page is a lag in updating, not an additional flaw.

What this means for university IT teams, open-source maintainers, and security teams

  • University IT teams: Treat exposed MONAI Label servers as high risk. Block untrusted network access to MONAI Label with firewall and routing rules; do not rely on obscured URLs. Turn on authentication (the server’s default is off), run the service under a limited account rather than root, and restrict writable folders.
  • Open-source maintainers: Merging a corrective change into source control is necessary but not sufficient. Users will need a packaged release; maintainers should accelerate shipping a versioned patch and clearly announce what builds include the October 19 fix so administrators can verify patched deployments.
  • Security operations and incident responders: If a server was reachable, check for unexpected Python startup files and unusual training jobs — the path the reported attack used — and assume that a writable code or settings location could be leveraged to run code as the service account.

Amazon’s reconstruction shows how quickly autonomous agents, when paired with exposed control infrastructure, can convert a logic bug into a significant compromise. For MONAI Label users the immediate choices are concrete: isolate the service, enable authentication, run it with minimal privileges, and verify any source builds include the October 19 correction. Whether the fix reaches wide deployment will determine if this incident remains an isolated university breach or a wider risk to lab deployments running the default, unauthenticated configuration.

Original story