AI Agents Found a Zero-Day in 11 Minutes: Lessons for Defenders
Amazon says attackers pointed an open-source AI agent platform called Cairn at university networks and it found and exploited an unknown flaw on its own. Here is what happened and how organisations should adjust.
In this article
For two years security teams have debated when AI would start hacking on its own. A report published by Amazon Threat Intelligence on 20 November 2026 suggests that moment has arrived, at least partly. Attackers gave a set of AI agents a one-line goal, “obtain shell access”, and the agents found a previously unknown vulnerability and used it to take root control of a server, with no human choosing each step. This piece explains what happened in plain language and what IT leads, universities and SMEs should change.
Key takeaways
- Attackers used Cairn, an open-source autonomous AI penetration-testing framework built for defenders, to run hacking operations with minimal human input.
- The agents went from mapping an application’s API to writing files on its server in about 11 minutes, using a zero-day path traversal flaw.
- Other attempts failed, but agents kept working for up to 16 hours straight, testing over 30,000 password combinations in one case.
- Basic hygiene, such as not exposing unauthenticated services and using strong credentials, stopped several attacks.
- Amazon’s main advice: rely more on high-confidence automated detection and response, with humans overseeing outcomes.

What Amazon found
On 10 September 2026 Amazon’s researchers came across a server running Cairn with no authentication. Because the operator had left it open, the platform’s history was visible, including project settings, the attacker’s prompts and the AI’s output. The prompts were written in Simplified Chinese and script timestamps used China Standard Time; Amazon attributes the activity to a China-linked actor with moderate confidence.
The campaign ran for about 48 hours from 8 September and targeted universities in East Asia outside mainland China. The operator supplied lists of IP addresses and variations of the goal “obtain shell access”. The agents then planned, acted and adjusted on their own.
The zero-day
Against a server running MONAI Label, an open-source tool that helps researchers label medical images using AI, the agents identified the software, mapped its API, found an upload function that accepted a filename, and tested it for path traversal. It was vulnerable. In a later session they used the flaw to plant a Python reverse shell in a file that Python loads automatically at start-up, then triggered a training function to launch Python, giving them a shell running as root.
Amazon says the flaw had no advisory, fix or prior public record at the time, so it assesses with high confidence that it was a true zero-day. The maintainers released a fix on 18 November after Amazon reported it. The operator then used the compromised server to tunnel into a second university network and pointed the agents at more machines.
The failures are just as instructive
Not everything worked. Against a Redis database left open without a password, the agents spent around 16 hours and 148 actions trying known takeover techniques, pivoting through a flaw in a neighbouring service and spraying passwords at PostgreSQL, SSH and others. None succeeded. Against a university GPU server, they built a list of about 80 real usernames from staff directories and papers, then tried more than 30,000 credential combinations across SSH, an admin dashboard, storage, VPN and a campus portal. They never got in.
Why this matters beyond universities
Three lessons stand out for any organisation, including Nigerian and African businesses that rely on small IT teams.
- Speed: Amazon notes that an agent that finds and exploits a flaw in minutes finishes faster than many security teams can triage a single alert manually.
- Stamina: agents do not sleep, take weekends or get bored. A 16-hour attack that would tire a human costs the attacker little.
- Access: Cairn is open source and easy to drive with plain-language goals. Gambit Security has documented a separate, financially motivated group using it against online retailers, and Google reported its use against a US aerospace organisation, so this is not limited to state actors.
Practical steps for defenders
Shrink what is exposed
- Inventory every internet-facing service, including research tools, AI apps and dashboards that staff or students set up themselves.
- Never expose databases like Redis, or admin panels, to the internet without authentication. Put them behind a VPN or firewall rules.
- Remove or patch niche open-source apps quickly; MONAI Label users should apply the 18 November fix.
Make credential attacks pointless
- Use multi-factor authentication on VPN, SSH and portals, and prefer SSH keys over passwords.
- Rate-limit and alert on repeated login failures. Remember attackers can harvest real usernames from your public website.
Automate the first response
- Pick a few high-confidence detections, such as a web server process spawning a shell or a sudden burst of failed logins, and have them trigger automatic blocking or isolation.
- Have humans review outcomes rather than approve every action. If you outsource security, ask your provider how fast their automated response is.
Use AI on defence too
The same class of tools can scan your own systems before attackers do. Drupal’s critical installer flaw fixed this week, for example, was found by an automated AI scanner. Running authorised scans of your own assets is now a reasonable step even for mid-sized organisations.
What you should do
- This week, list your internet-facing services and close anything that does not need to be public.
- Turn on MFA for every remote access path and block password-only SSH.
- Add the two IP addresses Amazon published (142.4.98.145 and 142.4.107.34, seen 8 to 10 September 2026) to your log searches.
- Agree with your IT team or provider on two or three alerts that should trigger automatic containment, and test them.
Sources
- AWS Security Blog: Chinese-nexus actor’s Cairn AI agents found a zero-day on their own
- BleepingComputer: Hackers use autonomous AI tool Cairn to find zero-day, take over server
Featured image: Photo: “Artificial Intelligence (AI) and Robotics exhibition at the Heinz Nixdorf MuseumsForum” by Sergei Magel/HNF, CC BY-SA 4.0, via Wikimedia Commons — source
Was this article helpful?
adminVerified expert
Cybersecurity Analyst
Federal Capital Territory (Abuja)
Credentials: MCPN, MNCS, IAENG | System Analyst · Web Developer · CEH · CC · CDST
Got expertise to share? Write for us →


