News
AI agents used a MONAI Label flaw to get root on a university server.
Amazon Threat Intelligence says suspected China-linked operators used autonomous AI agents to find a path-traversal flaw in MONAI Label and gain root on one university server. The flaw, CVE-2026-133968, affects version 0.8.5 and earlier when login is left off. There is still no fixed release to install.

Impact
- What was impacted
- MONAI Label, the open-source server research teams use to label medical scans and train imaging models. It is a separate program from the broader MONAI framework. CVE-2026-133968, scored 9.8, is a path-traversal flaw in the parts of the server that store uploaded images and labels. A file identifier supplied by the caller is appended to the storage folder and written without a check that the result stays inside that folder. Versions 0.8.5 and earlier are affected. Authentication is off by default, so any client that can reach the server can start the write. The server writes as the account it runs under. That does not mean a write anywhere on the machine. A writable code or settings location can still open a path to running code.
- Who was impacted
- Teams that run MONAI Label, especially where the server is reachable from a network they do not trust and login is still off. Amazon says the operators used the flaw for root on one university server, then later gave the agents a tunnel into a second university network running the same software. Amazon believes that second network was broken into the same way and has not confirmed it. The company ties the campaign to a China-linked actor with moderate confidence, from Simplified Chinese prompts, China Standard Time settings, and targets likely of interest to China. It did not name a group or show that a government was behind the work. The account is Amazon's reconstruction and has not been independently confirmed. It does not show that patient scans or records were taken, that clinical systems were touched, or that the attacker went past the university targets described.
- What you need to do
- There is no fixed release to install. The maintainers merged a source fix on October 19 that forces uploads to resolve inside the approved folders and rejects empty names or names that begin with a dot. Their advisory, published earlier this week, still lists no patched version and says the correction will arrive in a future release. The newest release listing remains 0.8.5, from 2024. Do not treat a source checkout as safe unless you confirm it includes that October fix. Block untrusted network access with firewall and routing rules, not only a hard-to-guess address. Turn authentication on. Run the server as a limited account rather than root, and restrict the folders it can write. If a server was reachable, look for unexpected Python files that run at startup and for unusual training jobs. That is the path the report says was used to get root. Do not send a traversal request to test the flaw.
Suspected China-linked operators used autonomous AI agents to find a previously unknown flaw in MONAI Label and gain root on a university server, Amazon Threat Intelligence said this week. The flaw lets a caller with no account write files outside the server's storage folder. It affects MONAI Label 0.8.5 and earlier in the default setup, which runs with no login. The maintainers have merged a fix in the source tree. There is still no fixed version to install.
MONAI Label is the server research teams use to label medical scans and train imaging models. It is not the broader MONAI framework. The weakness is a path traversal in the parts of the server that store uploaded images and labels. A file identifier from the caller is joined to the storage folder and written with no check that the path stays inside that folder. The write uses the server's own account, so it is not a write to every path on the machine. If that account can write code or settings, the flaw can lead to code execution. With authentication off by default, any client that can reach the server can try it.
Amazon reconstructed the intrusion from the operator's own Cairn control panel, which was sitting on the internet without a password. Cairn is an open-source framework that runs AI agents as automated testers. The open panel showed the agents' task settings, prompts, and run history. Over about 48 hours in September, the agents mapped the application and reached a file write in roughly 11 minutes. The operator set a goal of shell access and chose the targets. The agents picked the methods. Investigators link the campaign to a China-linked actor with moderate confidence, citing Simplified Chinese prompts, China Standard Time, and the choice of targets. They did not name a group or establish that a government directed it.
Root came later, in a separate session, not in those 11 minutes. The agents used the same flaw to plant a file that Python runs automatically at startup, then started a training feature. The new Python process loaded that file and connected back as root. That result was on one university server. Another MONAI Label system would match it only if it is run the same way and its service account can do the same things. In a later session the operator gave the agents a tunnel into a second university network running the same software. Amazon believes that network was entered the same way and has not confirmed the second compromise.
The maintainers confirmed the flaw and merged the source fix on October 19. Uploads must resolve inside the approved folders, and empty names or names that begin with a dot are rejected. The advisory published earlier this week still has no patched version and says the correction is coming in a future release. The newest listed release is still 0.8.5, from 2024. The flaw is CVE-2026-133968, published this week and scored 9.8. The project advisory still shows no CVE, which is that page lagging, not a second bug. Amazon's account has not been independently confirmed. It does not show stolen patient scans, touched clinical systems, or access beyond the university targets it describes. The campaign included failed operations. Root on one server is not the outcome for every deployment.
For your AI agent
Agent instructions — copy paste to Claude
You are helping me respond to CVE-2026-133968, a path-traversal flaw in MONAI Label 0.8.5 and earlier. Amazon Threat Intelligence says autonomous AI agents used it to gain root on one university server. There is no fixed release to install. Do not write or send a path-traversal request, a file upload, or a proof of concept. Do not start a training job or plant a Python startup file to test the bug. Ask before changing a firewall rule, turning authentication on, changing the service account, or replacing the program. Facts from the report: - MONAI Label is a server research teams use to label medical scans and train imaging models. It is a separate program from the broader MONAI framework. Do not flag the broader framework unless I also run MONAI Label. - The flaw is in the parts of the server that store uploaded images and labels. A caller-supplied file identifier is appended to the storage folder and written without a check that the result stays inside that folder. Versions 0.8.5 and earlier are affected. Authentication is off by default, so any client that can reach the server can start the write. - The server writes with the privileges of its own account. An attacker cannot write anywhere on the machine from this flaw alone. A writable code or settings location can open a path to running code. Do not describe every deployment as remote root. - Amazon reconstructed the case after finding the operator's Cairn control panel on the internet with no password. Cairn is an open-source framework that runs AI agents as automated penetration testers. The panel showed task settings, prompts, and run history. Do not go looking for other people's panels, and do not log into one. - Over about 48 hours in September, the agents mapped the interface and reached a file write in roughly 11 minutes. The operator set the goal of shell access and chose the targets. The agents chose the methods. Do not copy those prompts or methods. - Investigators tie the campaign to a China-linked actor with moderate confidence, from Simplified Chinese prompts, China Standard Time settings, and targets likely of interest to China. They did not name a group or establish that a government was behind it. Do not upgrade that to a named group. - Root came in a separate session, not in the 11 minutes. The agents planted a file Python runs automatically at startup, then triggered a training feature that started a new Python process. That process loaded the file and connected back as root. That was one university server. The same result elsewhere depends on how the server is run and what its service account can do. - A later session gave the agents a tunnel into a second university network running the same software. Amazon believes it was compromised the same way and has not confirmed it. Do not state the second compromise as fact. - The maintainers merged a source fix on October 19. Uploads must resolve inside the approved folders, and empty names or names that start with a dot are rejected. The advisory published earlier this week still lists no patched version and says the fix will arrive in a future release. The newest release listing remains 0.8.5, from 2024. A source tree is not safe unless it contains that October fix. - CVE-2026-133968 was published this week and scored 9.8. The project advisory still shows no CVE. That is a lag on the advisory page, not a second flaw. - The account is Amazon's reconstruction and has not been independently confirmed. It does not show that patient scans or records were taken, that clinical systems were touched, or that the attacker passed the university targets described. The campaign included failed operations. Do not invent data theft. Context: - What was impacted: MONAI Label 0.8.5 and earlier, default configuration with authentication off, and one university server where the reported chain reached root. - Who should care: teams that run MONAI Label, especially if the server is reachable from a network they do not trust. - What to do: there is no release to upgrade to. Cut off untrusted network access, turn authentication on, run as a limited account, restrict writable folders, and confirm any source build includes the October 19 fix. If the server was reachable, look for unexpected Python startup files and unusual training jobs. Do this in order. Ask before changing the server. Prefer read-only inventory first. Do not write or run an exploit or a proof of concept. 1. Inventory - Find whether I run MONAI Label. If I only run the broader MONAI framework, say this flaw is in MONAI Label and stop unless that server is also present. - Record the installed version. 0.8.5 and earlier are the affected releases named here. If the version is newer than 0.8.5, say this report does not call that newer release vulnerable, and do not invent a fixed version number. If I am running a source checkout, record the date or commit only if it is visible, and say safety depends on whether the October 19 fix is present. Do not assume a checkout is patched. - Record whether authentication is on, whether the service account is root, and whether the server is reachable from a network I do not trust. Read that from config or from what I show you. Do not probe the upload path. - If the server was reachable while still on 0.8.5 or earlier with authentication off, list Python startup files and recent training jobs if I can show you those directories. Describe anything unexpected. Do not delete files yet. 2. Contain - If the server is reachable from a network I do not trust, propose firewall or routing limits. A hard-to-guess address is not the control this report recommends. Ask before changing a rule. - Propose turning authentication on, running as a limited account, and restricting writable folders. Ask before each change. Do not claim those steps remove the flaw. They are the exposure limits in this report while no fixed release exists. 3. Remediate - Do not tell me to upgrade to a version number. This report says no patched release is available and that 0.8.5, from 2024, is still the newest listing. - If I build from source, say I should confirm the October 19 fix is in that tree before treating it as patched. Ask before replacing a deployment. 4. Verify - After a change I asked you to perform, read authentication, the service account, and the network exposure again. - If I did not ask you to change anything, report the inventory only. 5. Report back - Whether MONAI Label is present, its version, whether authentication is off, whether it runs as root, whether untrusted networks can reach it, what you changed, and that root in the report was one university server, not the result for every install. Do not invent CVEs, release numbers, patient-data theft, or a named threat group. If a check needs access I have not given you, say what you need and stop.