Autonomous AI agents found and exploited a zero-day vulnerability on their own, taking full control of a university server in about 11 minutes. Amazon disclosed the campaign on November 20, calling it the first documented case of agents discovering a flaw and using it in the wild.
The agents were running Cairn, an open-source framework that coordinates AI agents for security testing. Their operator gave them a single goal, ‘obtain shell access’, and a list of university targets, then let the agents plan, execute and adjust their own approach.
Eleven Minutes to Root
Against a medical imaging application called MONAI Label, the agents mapped the interface, found an upload endpoint with a filename parameter, and tested it for path traversal. The parameter was vulnerable, and the agents went from mapping the interface to writing files on the server in about 11 minutes.
In a later session, the agents used the same flaw to write a Python reverse shell into a file the interpreter loads at startup, then triggered a training endpoint to spawn a new process. The process imported the payload and opened a reverse shell running as root.
Amazon says it has high confidence the flaw was a zero-day, because no fix, issue, advisory or prior mention existed in vulnerability databases at the time. The project’s maintainers have since merged a fix and published a security advisory, but no patched release is available yet.
A Full Day Without Direction
The zero-day was the success in a campaign that ran about 48 hours from September 8, against universities in East Asia outside mainland China. The agents also spent 16 hours and 148 actions on an open Redis server, and six hours spraying 30,000 credential combinations at a GPU server.
Neither of those attempts succeeded, yet the agents kept adapting for nearly a full day without human direction. Amazon found the campaign because the operator left the Cairn server publicly accessible, exposing its prompts, scripts and outputs to anyone on the internet.
Amazon assesses with moderate confidence that the actor is China-linked, based on Simplified Chinese prompts, China Standard Time timestamps, and targets likely of strategic interest to China. The attribution is an assessment, not proof of state direction, and the account itself is Amazon’s reconstruction from the exposed server.
Our Read
This is the moment the agent threat stopped being theoretical. Until now, the debate was whether agents could help attackers; this campaign shows agents doing the whole job, from finding the flaw to opening the shell, while the operator watched.
The speed is what defenders should fear. Eleven minutes from mapping to file write is faster than most security teams can triage an alert, and the agents that failed ran for 16 hours straight without breaks. Persistence at machine speed is a new kind of adversary.
Why It Matters
The same framework is already being used by criminals, not just spies. Amazon notes that a financially motivated actor has used Cairn against online retailers, and researchers documented the same operator stealing payment records with other AI tools, from a major airline and a hospitality company.
For defenders, the lesson is automation. Amazon says organisations should prioritise detections that trigger automated response, with humans overseeing outcomes rather than reviewing every action. An agent that works for 16 hours without sleep cannot be matched by a team that works in shifts.
Background
The campaign is not the first time Cairn has shown up in the wild. Gambit Security documented a financially motivated actor using the same framework against online retailers, and Amazon says the tool is now being adopted in both espionage and criminal operations.
Source: AWS on the Cairn zero-day and The Register on the campaign


0 Comments