Subscribe
Cybersecurity

Autonomous AI agents found a zero-day on their own — 11 minutes to break in, then root

Sunday, November 22, 2026 · 6 min read

Amazon Threat Intelligence disclosed on Nov 20 that autonomous AI agents using the Cairn framework found a zero-day in the MONAI Label medical imaging app, wrote files anywhere on the server within 11 minutes, and took root shell in a later session — with China-nexus attribution at Amazon's moderate confidence.

Bearish$AMZN$NVDA

📊 Scorecard: our Bearish call on $AMZN is scored from $246.82, the last close before it ran. Its first result comes after the next close.

Autonomous AI agents found a zero-day on their own — 11 minutes to break in, then root

The News

Amazon Threat Intelligence disclosed on Nov 20 that a suspected China-nexus threat actor used Cairn, an open-source autonomous AI penetration-testing framework, to discover a zero-day vulnerability in a medical imaging application and exploit it to take complete control of a server. The operators gave the AI agents a single instruction — "obtain shell access" — and left them to discover and exploit the flaw independently. The campaign ran Sept 8–10, 2026, targeting mainly East Asian educational institutions outside China.

The agents found a path traversal zero-day in MONAI Label, an open-source medical imaging application used to label MRI and CT studies, from the MONAI consortium co-founded by NVIDIA. Within 11 minutes they went from mapping the application's API to writing files to arbitrary locations on the server. In a subsequent session they wrote a Python reverse shell into sitecustomize.py — a file Python loads automatically at startup — then triggered a training API endpoint to establish root access. The flaw, CVE-2026-133968 (CVSS 9.8), affects MONAI Label 0.8.5 and earlier, which ships with authentication disabled by default; the fix was merged Oct 19 (PR #1915) and the GitHub advisory GHSA-qc22-rmgc-vg94 was published Nov 18, but researcher write-ups as of Nov 21 said no packaged release yet contained the fix.

Amazon found the campaign because the operator left its Cairn instance publicly accessible without authentication, exposing its prompts and execution history. After the zero-day, the operator gave the agents a proxy into a second university network running the same vulnerable software; Amazon believes that network was broken into the same way but has not confirmed the second compromise. The agents were not always successful: a 16-hour, 148-action effort to take over a Redis server failed, and over 30,000 credential combinations tried over six hours against a university GPU server failed to brute-force entry. Amazon attributes the campaign to a China-nexus actor with moderate confidence, based on the Simplified Chinese prompts, China Standard Time settings in the scripts, and target selection.

One instruction — obtain shell access — and the AI agents found a zero-day and wrote files anywhere on the server in 11 minutes, then took root.

Why It Matters

This matters because it is the clearest case yet of AI agents finding a zero-day and exploiting it in the wild with almost no human guidance. Previous AI hacking stories involved humans using AI as a helper; here the agents did the discovery, the exploitation and the persistence themselves, from one short instruction. The 11-minute timeline is the headline number, but the real shift is the labour model: an operator can now run many such campaigns in parallel, and the agents do not get tired or make typos.

The target makes it worse. MONAI Label is used to label MRI and CT studies — medical imaging, in universities and hospitals — and it ships with authentication disabled by default, so anyone who can reach it can use it. The fix was merged on Oct 19 and the advisory published Nov 18, yet as of Nov 21 no packaged release contained the fix, which means vulnerable installations are still out there waiting to be updated by hand. A medical app with no login and no patched release is a sitting target.

The failures are worth noting too, because they show the limits. A 16-hour, 148-action effort to take over a Redis server failed, and more than 30,000 credential combinations tried over six hours failed to brute-force a university GPU server. Cairn is not magic — it is a fast, tireless junior hacker that still hits walls. But the walls it does not hit are the ones that matter: a zero-day nobody knew about, found and used in minutes, against software that guards medical images.

Analyst's Take

Our call is Bearish. Autonomous AI agents that can find a zero-day, write a backdoor and take root access from a single instruction change the economics of hacking: the skill that once took a trained researcher now takes a prompt. The target here was a medical imaging app used to label MRI and CT studies, running with authentication disabled by default — and as of Nov 21, no packaged release contained the fix.

What to watch next is whether a packaged MONAI Label release ships the fix, and whether other Cairn-driven campaigns surface. Amazon found this one only because the operator left its own Cairn instance publicly accessible without authentication — a careless mistake that exposed its prompts and execution history. Other operators may not be so careless. The unconfirmed second-network compromise is also worth watching: if Amazon confirms it, the campaign's reach doubles.

For investors, the read-throughs are Amazon and NVIDIA — and both carry mixed calls here, not bearish ones. Amazon is the discloser, not the breached party: its threat-intelligence team found and published the campaign. NVIDIA co-founded the MONAI consortium behind the vulnerable app, but it did not write or ship MONAI Label and was not breached. Neither company is the victim, so neither deserves the story's bearish verdict — but both are now attached to the hardest question in security: what happens when the attacker is an AI agent that never sleeps. Reader poll: Should AI agents be allowed to hunt zero-days — yes or no? Vote: Yes, with limits, No, too risky, Only in labs.

Are you bullish or bearish on this story?

By The Numbers

11 minutesFrom mapping the app's API to writing files anywhere on the server
9.8CVSS score for CVE-2026-133968
Sept 8–10Campaign dates, targeting East Asian educational institutions
30,000+Credential combinations tried over six hours — and failed
16 hours148-action effort to take over a Redis server — also failed
Nov 18GitHub advisory published; no packaged release yet had the fix

FAQ

What is a zero-day?

A zero-day is a security flaw that nobody knew about before it was used — not the software maker, not the defenders. Because no fix exists yet, the attacker has a free run until the flaw is discovered and patched. This one, CVE-2026-133968, was found by AI agents, not by a human researcher.

What is path traversal?

Path traversal is a flaw that lets an attacker escape the folder an app is supposed to stay inside and reach files elsewhere on the server. Here the agents used it to write files to arbitrary locations on the server — the step that turned a flaw in a medical imaging app into full control of the machine.

What is a reverse shell?

A reverse shell is a way for an attacker to take control of a server from a distance: a small program on the server opens a connection back to the attacker's machine, giving them a command line as if they were sitting at the server. The Cairn agents wrote one into a file Python loads automatically at startup.

What does moderate confidence mean?

Moderate confidence is an intelligence term meaning the evidence points in a direction but is not conclusive. Amazon bases its China-nexus attribution on Simplified Chinese prompts, China Standard Time settings in the scripts, and the choice of targets — suggestive, not proof.

👁

Source: BleepingComputer

Get the daily briefing. AI news with a markets lens — tickers, bullish/bearish reads, zero hype. Subscribe free on Substack

More on $AMZN

All $AMZN coverage →

More from AI Agent Update