DKDKCISSP
DKCISSP / BLOG

AI Agents Found a Zero-Day in Minutes. Is Your Team Faster?

An attacker told AI agents to obtain shell access, and they found and exploited an unknown flaw in a medical-imaging server on their own, in about 11 minutes. How the Cairn agents did it, and what machine-speed attacks mean for defenders.

By Deepak Kumar, CISSP1 min readOpinion

For years we comforted ourselves with one idea: finding a zero-day is hard. It takes skilled people, time and money, so only a few serious groups can do it. This week that comfort took a hit.

Amazon’s threat intelligence team described an attacker who simply told a set of AI agents to “obtain shell access” on a server running a medical-imaging tool. The agents mapped the application, found a flaw nobody knew about, and got root. The human picked the target. The machines did the rest.

The message is simple: attack speed is no longer limited by how many skilled people an attacker has. And this was not the only AI attack tool in the news this week; I cover ARTEX, the FBI’s other warning, in its own post.

Find the flaw11 minroughly, from mapping the app to writing files on the server
Groundwork48 hoursabout how long the agents spent mapping the API first
Severity9.8CVSS for the MONAI Label flaw, CVE-2026-133968
Fix statusNo releasea source fix exists, but no installable fixed version yet
01 · How the AI agents got root

One instruction, five moves, no exploit written by a human

Reconstructed by Amazon from the attacker’s own control panel, which was left open on the internet without a password.

THE HUMAN Sets one goal “Obtain shell access” and picks the target. Prompts were in Simplified Chinese AGENTS · ABOUT 48 HOURS Map the app Explore the MONAI Label API and find an upload endpoint AGENTS · 11 MINUTES Find a zero-day Path traversal bug: write a file anywhere. No login: auth is off by default AGENTS · LATER SESSION Plant and trigger Drop a Python startup file, then start a training job that loads it RESULT Root shell on a university server. A second network was reached the same way (not confirmed) The flaw is CVE-2026-133968 (CVSS 9.8). A source fix was merged on 19 October, but there is still no installable fixed release.

In plain words: MONAI Label lets researchers upload medical images and labels. The upload code trusted the file name it was given, so a name with the right “../” parts could write a file anywhere the server could write. Because login was switched off by default, anyone who could reach the server could do it. The agents found this on their own, then used it to drop a small Python file that ran with full root rights the next time a training job started.

The irony is how we know. The attacker left the AI control panel open on the internet without a password, so Amazon could read the agents’ instructions and history. The attacker automated the hacking but not their own security. Do not count on that luck next time.

02 · Not a one-off

The same week, the FBI warned about ARTEX

Cairn is not alone. On 13 November the FBI warned that another AI framework, ARTEX, has been used since August to automate whole intrusions, some reaching privileged access in under three hours. It deserves its own deep dive, so I wrote one: ARTEX: the pentest tool that became everyone’s attack team.

03 · Race yourself

Would you finish before the machine does?

Put in your own numbers. The attacker side uses only the two figures reported this week.

How long until someone looks at a new high-severity alert?
minutes
How long to patch a critical flaw on an internet-facing system?
days
Can your team cut off a host automatically, without waiting for a human?

Why compare these numbers? Because defence is a race. If an attacker can get from first look to a working foothold faster than your team reads the alert, detection alone will not save you. The attacker figures here are single reported cases, not averages, so treat the result as a wake-up call rather than a precise measurement.

04 · What changes for defenders

Four assumptions this week broke

Most security programmes quietly rely on a few comfortable assumptions. AI agents do not make attacks smarter so much as tireless and fast, and that breaks the assumptions that depended on attackers being busy, bored or short of people.

Broken

“Nobody will look at that old app”

Agents read every endpoint, every time. Obscure and forgotten systems are exactly where the FBI says ARTEX does best.

Broken

“Default settings are fine for now”

MONAI Label shipped with authentication off. One default setting turned a file-write bug into root.

Broken

“We patch within the month”

An agent can find and use a flaw in minutes. A monthly cycle is now a very long open window.

Broken

“A human will review the alert”

Amazon’s own advice: let machines contain the obvious cases automatically, and keep humans for judgement.

05 · Monday morning

What I would put on the board on Monday

Automate the first five minutes

Pre-approved playbooks that isolate a host or disable an account on high-confidence alerts.

Use AI on your own estate first

Point the same kind of tooling at your own apps, with permission and scope, before someone else does.

Rate-limit and watch for bursts

The FBI’s advice: monitor for floods of automated requests, sudden privilege changes and large transfers.

The attacker’s team just got a lot bigger. Ours has to get faster.

AI did not invent a new kind of attack this week. It took the attacks we already know and removed the waiting. Our defences were built around the time it takes a skilled human to get through a network. That time is shrinking, so the basics have to be done before the alert, not after it.

Related on DKCISSP: When Cyberattacks Start Thinking: AI, Autonomous Threats, and Connected Systems

Sources

Facts as reported up to 21 November 2026. The Cairn intrusion is Amazon’s account and has not been independently confirmed.

  • AI Security
  • Zero-Day
  • Threat Intelligence
  • Opinion
← Back to Blog