Amazon says AI agents cracked a university server in 11 minutes
Amazon Threat Intelligence caught a China-linked operator steering AI agents through university servers in September, and its message to defenders is to stop approving every move by hand. One zero-day fell in about 11 minutes, while a Redis box held out for 16 hours. The flaw, in MONAI Label, is now patched and rated critical.
- Original source: cryptopolitan.com.
- The source published it on Nov 20, 2026 at 11:34 PM.
- No second outlet has reported it yet. Treat it as a single-source claim.
Amazon Threat Intelligence detected the campaign on Sept 10, 2026, and the company's advice is blunt: stop clearing every move by hand. CJ Moses, chief information security officer at Amazon Integrated Security, says defenders should wire high-confidence detections straight into automatic responses and let humans watch the outcome. An agent that finds and exploits a zero-day in minutes, he wrote, gets done faster than many security operations centers can manually triage an alert.
The operator behind it was linked to China with moderate confidence, based on Chinese Simplified prompts, scripts set to China Standard Time and targets of likely strategic value to Beijing. The giveaway was carelessness: the operator left a Cairn server running with no login, exposing its prompts, settings and AI output to anyone who looked. Cairn is an open-source platform that walks several agents through penetration tests. Logs showed IP addresses tied to East Asian universities outside mainland China hit for about 48 hours from Sept 8, with one target list and variations of a single order: obtain shell access.
A server running MONAI Label, an open-source medical imaging AI, went first. The agents mapped its API, found an upload endpoint with a filename parameter and probed it for path traversal, and roughly 11 minutes after mapping the API they could write files anywhere on the machine. Another session dropped a Python reverse shell into sitecustomize.py and seized root. Amazon rated it a zero-day with high confidence, since there was no fix, advisory or earlier mention; it alerted the maintainers, the code got patched, and a Nov 18 GitHub advisory calls the bug critical and credits the AWS report.
From there the agents were handed a proxy into a second university's network where a server ran the same vulnerable software. Amazon believes the operator used the zero-day to compromise that machine and open the tunnel.
Against a university Redis server left open with no password, the agents logged 148 consecutive actions over roughly 16 hours, abusing Redis replication to inject rogue code, escaping the scripting sandbox, rewriting config files, corrupting memory and running a Log4Shell exploit on another service on the host. They then mapped the internal Docker network through a server-side request forgery hole in a Qdrant database. Password spraying at SSH, PostgreSQL and other services followed. None of it worked.
A GPU server also held, surviving more than six hours and over 30,000 credential combinations without a single success, though the agents compiled around 80 real usernames from faculty directories, lab sites, archived pages and papers. Moses wrote that they ran continuously for 16 hours without breaks, sleep, or working-hour constraints. AWS alerted the victims through a government partner because none of them were customers. Cairn is surfacing elsewhere too: Gambit Security separately documented a financially motivated group using it against online retailers, and Google researchers said on Nov 17 that they found it in intrusions at American aerospace companies.
TopicsAmazon · AWS · Cairn · MONAI Label · Qdrant · CJ Moses · Google · Gambit Security
Source: cryptopolitan.com
Written by AI from verified reporting.
More stories
- Tech & AI cryptobriefing.com 2 viewsGoogle Axes the Search Button on Its Homepage to Push Gemini
- Tech & AI blockchain.news 1 viewFBI Flags ARTEX: Hacking AI Turned Loose on Real Targets Since August
- Tech & AI dailymail.com 2 sources 1 viewAnthropic staffer claims Claude just cracked the Riemann hypothesis
- Tech & AI cryptopolitan.com 1 viewAnthropic co-founder joins Musk and the Dells as a Trump Accounts donor

