{"slug": "china-linked-hackers-ai-agents-found-a-brand-new-flaw-on-their-own-and-seized-a", "title": "China-linked hackers’ AI agents found a brand-new flaw on their own and seized a university server, Amazon says", "summary": "Amazon Threat Intelligence reported on Friday that AI agents coordinated by the open-source Cairn framework autonomously discovered a previously unknown zero-day vulnerability in MONAI Label, an open-source medical imaging AI tool, and used it to gain root shell access to a university server in about 11 minutes, from mapping the target's software to writing files on it. Amazon said it found the operation on September 10 after spotting a Cairn server with no password, and traced roughly 48 hours of activity from September 8 against internet addresses at universities in East Asia outside mainland China, attributing the campaign to a Chinese-nexus actor with moderate confidence based on Simplified Chinese prompts and China Standard Time scripts. Amazon said it has high confidence the flaw was a zero-day because no fix, bug report, advisory or earlier mention existed, and most of the campaign's other attempts failed, including about 16 hours and 148 actions against an open Redis database and more than six hours against a university GPU server.", "body_md": "A US Air Force radiologist reviews CT scans at Balad Air Base, Iraq, in February 2008. Illustrative: MONAI Label is used to label scans like these to train medical AI. Image: [Senior Airman Julianne Showalter, US Air Force](https://commons.wikimedia.org/wiki/File:X-ray_Vision_DVIDS78078.jpg) / Wikimedia Commons, Public domain, cropped\n\nHackers Amazon links to China gave a team of AI agents one goal, “obtain shell access”, and the agents found a security hole nobody knew about and used it to take over a university server, Amazon’s threat intelligence team [said on Friday](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/). In about 11 minutes they went from mapping the target’s software to writing files on it. The flaw was in MONAI Label, an open-source medical imaging AI tool.\n\n## An open Cairn panel gave the game away\n\nAmazon Threat Intelligence says it found the operation on September 10, when it came across a server running Cairn with no password on it. Cairn is an open-source framework built for defenders: given a starting point and a goal in plain language, it coordinates several AI agents to find a way there. Because it was public, Amazon could read the attacker’s settings, prompts and the agents’ output.\n\nThe records showed a campaign of about 48 hours from September 8 against internet addresses belonging to universities in East Asia outside mainland China. Amazon names neither the universities nor the AI models behind the agents. It ties the work to a Chinese-nexus actor with moderate confidence, citing prompts written in Simplified Chinese, scripts set to China Standard Time and targets “likely of strategic interest to China”.\n\nCairn has turned up in attacks before. In October one hacker’s exposed Cairn panels showed [AI agents loose on 1,200 websites](https://madrobot.blog/2026/10/19/hacker-ai-agents-cairn-1200-websites-deere-sharkninja-norforge/), and [Gambit Security reported](https://gambit.security/blog-posts/autonomous-ai-agents-online-retailers-25-a-company) a financially motivated operator using it with two other open-source tools to hit online retailers, launching 105 attack projects between September 10 and 15 and compromising at least 27 companies. Google has also [seen it used against US aerospace firms](https://madrobot.blog/2026/11/16/google-hackers-ai-patches-exploits-no-human/).\n\n## Eleven minutes to a brand-new flaw\n\nAgainst a server running MONAI Label, which helps researchers label CT and MRI scans to train AI models, the agents worked out what software it was, mapped its API and found an upload feature that took a filename. They tested whether a crafted filename could write a file outside the folder it was meant for. It could. That took about 11 minutes.\n\nIn a later session the agents used the same hole to plant a Python reverse shell in sitecustomize.py, a file Python loads every time it starts, then called the tool’s training feature to launch a new Python process. That gave them a remote shell running as root: full control of the server.\n\nAmazon says it has high confidence this was a **zero-day**, a flaw unknown to the people who make the software, because there was no fix, bug report, advisory or earlier mention of it anywhere. A human then stepped in. The operator handed the agents a proxy tunnelling into a second university’s network through a server running the same software, which Amazon infers had been hacked the same way, and told them to go after other machines there.\n\n## “Adapting for nearly a full day”\n\nMost of the campaign failed. Against an open Redis database at one university, the agents spent about 16 hours and 148 actions working through known ways to take over the server, tried Log4Shell on another service, used a flaw in a Qdrant database to map the internal network and sprayed passwords at PostgreSQL, SSH and other services. None of it worked.\n\nAnother run of more than six hours went after a university GPU server. The agents built a list of about 80 real usernames from staff directories, lab websites, archived pages and academic papers, then tried more than 30,000 username and password combinations across the server, an admin dashboard, a storage device, a VPN and the campus portal. They never got in. Amazon’s security chief wrote:\n\nNeither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction.\n\nCJ Moses, chief information security officer, Amazon\n\n## Fixed after Amazon’s warning\n\nAmazon told MONAI Label’s maintainers, who [fixed the code](https://github.com/Project-MONAI/MONAILabel/pull/1915) and published a [GitHub security advisory](https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94). The flaw is now [CVE-2026-133968](https://nvd.nist.gov/vuln/detail/CVE-2026-133968), rated 9.8 out of 10, covering MONAI Label up to version 0.8.5, “as exploited in the wild starting in September 2026”. MONAI Label is part of Project MONAI, the open-source medical imaging AI project Nvidia started with King’s College London.\n\nThe universities were not AWS customers, so Amazon says it worked with a government partner to warn them. Moses says defenders now need automated responses to keep up:\n\nAn agent that discovers and exploits a zero-day in minutes completes its activity faster than many security operations centers can manually triage an alert.\n\nCJ Moses, chief information security officer, Amazon\n\n## Why it matters\n\nAI hacking agents have mostly been caught working through known flaws and stolen or guessed passwords. Here, agents found a flaw nobody had reported, in a niche research tool, in minutes, with no one telling them how. Any server left open to the internet is now worth an agent’s time.\n\n*Sources: [Amazon Threat Intelligence, “Chinese-nexus actor’s Cairn AI agents found a zero-day on their own” (AWS Security Blog)](https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/); [MONAI Label security advisory GHSA-qc22-rmgc-vg94](https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94); [NVD, CVE-2026-133968](https://nvd.nist.gov/vuln/detail/CVE-2026-133968); [Gambit Security](https://gambit.security/blog-posts/autonomous-ai-agents-online-retailers-25-a-company).*", "url": "https://wpnews.pro/news/china-linked-hackers-ai-agents-found-a-brand-new-flaw-on-their-own-and-seized-a", "canonical_source": "https://madrobot.blog/2026/11/21/amazon-cairn-ai-agents-zero-day-university/", "published_at": "2026-11-21 10:12:00+00:00", "updated_at": "2026-11-21 10:41:41.888414+00:00", "lang": "en", "topics": ["ai-agents", "ai-safety", "ai-policy"], "entities": ["Amazon", "Amazon Threat Intelligence", "Cairn", "MONAI Label", "Gambit Security", "Google", "Redis", "Qdrant"], "also_reported_by": [], "takeaway": null, "recommendation": null, "alternates": {"html": "https://wpnews.pro/news/china-linked-hackers-ai-agents-found-a-brand-new-flaw-on-their-own-and-seized-a", "markdown": "https://wpnews.pro/news/china-linked-hackers-ai-agents-found-a-brand-new-flaw-on-their-own-and-seized-a.md", "text": "https://wpnews.pro/news/china-linked-hackers-ai-agents-found-a-brand-new-flaw-on-their-own-and-seized-a.txt", "jsonld": "https://wpnews.pro/news/china-linked-hackers-ai-agents-found-a-brand-new-flaw-on-their-own-and-seized-a.jsonld"}}