SecureCyberVM
CVE Lookup Daily Threat Briefing Exploited CVEs Ransomware Zero-day
SecureCyberVM
Get the Patch Brief

Amazon: China-nexus operator used Cairn AI agents to find and exploit MONAI Label zero-day

Amazon Threat Intelligence says a suspected China-nexus operator used the Cairn AI pentesting framework to discover a path traversal zero-day in MONAI Label and breach two East Asian universities. A fix was released on 18 November.

Amazon: China-nexus operator used Cairn AI agents to find and exploit MONAI Label zero-day

Amazon Threat Intelligence says a suspected China-nexus operator used Cairn, an open-source autonomous AI penetration-testing framework, to find and exploit a previously unknown path traversal flaw in MONAI Label, an open-source medical imaging annotation application, and gain root access to a university server. According to BleepingComputer and The Record, the same weakness was then used to reach a second university running the vulnerable software.

The findings, published by Amazon on 20 November 2026, describe what the company assesses with high confidence to be a zero-day discovered by AI agents with little step-by-step direction from a human. MONAI Label maintainers released a security update on 18 November, according to BleepingComputer. No CVE identifier has been reported for the flaw; the only public identifier cited is GitHub advisory GHSA-qc22-rmgc-vg94.

What happened

According to The Record, Amazon began investigating on 10 September after finding a publicly accessible server hosting a Cairn instance whose prompt history was visible. BleepingComputer places the observed attack activity between 8 and 10 September 2026. The operator supplied target IP addresses and a plain-language objective, reported by both outlets as an instruction to obtain shell access. The agents then planned, executed and adjusted their approach on their own.

Not every attempt succeeded. Both outlets describe a failed effort lasting about 16 hours and involving 148 actions, although they differ on what it targeted. The Record also reports a failed brute-force attempt against a university GPU server, in which more than 30,000 credential combinations tied to 80 real usernames were tried over roughly six hours.

The successful intrusion was faster. BleepingComputer reports that the agents went from mapping the MONAI Label API to exploiting the path traversal flaw in about 11 minutes. The flaw allowed files to be written to arbitrary locations on the server. In a later session the agents planted a reverse shell in a Python file that loads automatically at interpreter start-up, then triggered an API endpoint to obtain root access. The operator subsequently gave the agents a proxy into a second university network running the same software.

Amazon says it worked with a government partner to notify the affected organisations, according to The Record. Amazon contacted the MONAI Label maintainers, who fixed the bug.

Who is affected

The victims were universities in East Asia; BleepingComputer describes them as mainly educational institutions outside China. Amazon declined to name the institutions or the country. Any organisation running an unpatched, network-exposed MONAI Label server is potentially exposed to the same flaw. The fixed version number has not been disclosed in the reporting reviewed.

Severity and exploitation

The flaw was exploited before a fix existed, and exploitation led to root-level code execution, which places it in the high-severity range for affected deployments. Amazon attributes the activity to a China-nexus actor with moderate confidence, citing Simplified Chinese prompts, timestamps set to China Standard Time and the choice of targets. No named group has been identified.

BleepingComputer published two IP addresses linked to the activity between 8 and 10 September: 142[.]4.98.145 and 142[.]4.107.34. No detection signatures were included in the reporting reviewed.

Amazon’s CISO, CJ Moses, warned that the agent completed its work faster than many security operations centres “can manually triage an alert”, according to The Record.

What defenders should do now

  1. Identify any MONAI Label deployments, including research and lab servers outside central IT management, and apply the maintainers’ security update referenced in advisory GHSA-qc22-rmgc-vg94.
  2. Remove MONAI Label and similar research AI services from direct internet exposure; place them behind VPN or authenticated reverse proxies.
  3. Check affected hosts for unexpected changes to Python start-up files such as sitecustomize.py, unfamiliar files written outside the application directory, and unexplained outbound connections.
  4. Search firewall, proxy and application logs for the two IP addresses published by BleepingComputer, keeping in mind the activity dates back to September.
  5. Alert on high-volume credential attempts against GPU and research servers, and enforce MFA or key-based access where possible.
  6. Review detection and response processes for machine-speed intrusions, prioritising automated containment over manual triage, as Amazon recommends.

SCVM analysis

This is one of the more concrete public accounts so far of an AI agent framework discovering and weaponising an unpublished flaw in a live intrusion, rather than in a lab. The target choice matters: niche, research-focused AI applications are often deployed by academic teams, sit outside patch management and are rarely monitored. Those are the conditions in which an agent that can test an API for hours without fatigue will find something.

Three details stand out for defenders. The successful exploit took minutes, not days. The failed attempts were long and noisy, with thousands of login attempts and well over a hundred actions, which means detection opportunities existed before the eventual breach. And the tooling is open source, so the same capability is available to any actor. The relevant control is less about the AI itself than about exposure: unpatched, internet-facing research services are now a faster path to compromise than before.

Sources

About this article: produced by the SCVM Newsroom with AI assistance from the public sources linked above, enriched with NVD, FIRST EPSS and CISA KEV data, and checked by automated quality controls under our editorial policy. Spotted an error? Request a correction.

SecureCyberVM

Install SecureCyberVM

Get the Daily Threat Briefing and exploited-CVE alerts one tap away, full screen, like an app.