Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 5 min read

The Eleven-Minute Zero-Day: Autonomous AI Agents Found and Exploited a Flaw No Human Had Reported

On November 20, Amazon's Threat Intelligence team published a report that reads like a milestone the security industry has been expecting with some dread. A threat actor gave autonomous AI agents a plain-language objecti

On November 20, Amazon's Threat Intelligence team published a report that reads like a milestone the security industry has been expecting with some dread. A threat actor gave autonomous AI agents a plain-language objective - "obtain shell access" - against a list of university IP addresses. The agents planned, executed and adapted on their own, and against one target they discovered a previously unknown vulnerability in an open-source application and exploited it to take full control of a server. Amazon assesses with high confidence the flaw was a genuine zero-day: at the time of the attack, no fix, issue, advisory or prior mention existed anywhere for it.

The disclosure matters not because the agents were unstoppable. Most of what they tried failed, sometimes spectacularly. It matters because of the one thing that worked, and how fast it worked.

How Amazon saw the whole campaign

The investigation exists because of an operational mistake by the attackers. On September 10, Amazon's team found a server hosting a public instance of Cairn - an open-source autonomous AI penetration-testing framework - running without authentication. The platform's entire operational history was visible: project configurations, the operator's prompts, the agents' output. The prompts were written in Simplified Chinese, script timestamps were set to China Standard Time, and the targets were universities in East Asia outside mainland China, over a roughly 48-hour window beginning September 8. Amazon attributes the campaign to a Chinese-nexus actor with moderate confidence on that basis.

Cairn itself is a legitimate tool, built as a research and problem-solving engine for defenders: give it a starting point and an objective in plain language, and it coordinates multiple AI agents to find a path to the goal. The same ease of use is what makes it attractive to attackers. In this campaign, the human supplied target IPs and the objective; the agents chose the steps.

The failures are the honest part

Before the breakthrough, the agents lost twice, and Amazon's report details both losses with unusual candor.

Against a Redis instance exposed without authentication at one university, the agents ran for roughly 16 hours and 148 sequential actions. They worked through the known playbook for open Redis servers - abusing replication to load unauthorized code, escaping the scripting sandbox, writing configuration files, corrupting memory - then attempted Log4Shell injection against a co-located service, mapped the internal Docker network through a server-side request forgery flaw in a neighboring Qdrant instance, and sprayed credentials against PostgreSQL, SSH and other services. Nothing worked.

Against a university GPU cluster, a separate six-hour operation first assembled about 80 real usernames from the university's public footprint - faculty directories, lab websites, archived pages, academic papers - then tried more than 30,000 credential combinations across SSH, an admin dashboard, a network storage device, a VPN and the campus portal. It never found a working login.

Neither operation succeeded. As Amazon's report puts it, together they show the agents kept adapting for nearly a full day without human direction - no breaks, no sleep, no working-hour constraints.

Eleven minutes to arbitrary file write

The third target ran MONAI Label, an open-source medical imaging application that uses AI to help researchers annotate scans. There the agents identified the framework, mapped the API, and found an upload endpoint with a caller-controlled filename parameter. They tested it for path traversal. It was vulnerable. About 11 minutes elapsed between mapping the application's API and writing files to arbitrary paths on the server.

The mechanics, per the maintainers' security advisory published November 18 (GHSA-qc22-rmgc-vg94), are as simple as they are severe. The PUT /datastore/image endpoint accepts an image query parameter used as the image ID without validation; the datastore code appends the file extension, joins it onto the image directory and writes the upload with shutil.copy(), never checking that the resolved path stays inside the datastore. A value like ../../../tmp/x writes anywhere the process can write. The label endpoint builds its destination the same way. The weakness is CWE-22, rated critical - and MONAI Label ships with authentication disabled by default, so any network client that can reach the server can exploit the default configuration.

In a later session the agents turned file write into code execution: they wrote a Python reverse shell into sitecustomize.py, a file the Python interpreter imports automatically at startup, then triggered a training API endpoint to spawn a fresh Python process. The new process imported the payload and opened a reverse shell running as root.

Then the human stepped back in. The operator handed the agents a proxy tunneling into a second university network - Amazon infers the actor had already compromised that server with the same zero-day - and directed them against other machines inside it.

The patch gap

Amazon notified the MONAI Label maintainers, who fixed the code and published the advisory on November 18. The fix (pull request #1915) resolves the destination path and rejects anything outside the datastore directories, and the image endpoint now reduces the ID to its basename and rejects empty or dot-prefixed IDs with HTTP 400. But the advisory's own "Patched versions" field reads: None. The fix is merged in the repository and will ship in the next release; as of November 20, there is no tagged release containing it. Every version up to and including 0.8.5 is affected, and the flaw is tracked as CVE-2026-133968.

Until a release ships, the advisory's workarounds are the whole defense: enable authentication (MONAI_LABEL_AUTH_ENABLE=True) and grant the ANNOTATOR role carefully, do not expose the server to untrusted networks, and run it as an unprivileged user with write access limited to the datastore directory.

Not an isolated case

Amazon's report places the campaign in a pattern. Gambit Security previously documented a financially motivated actor using Cairn in at least 105 attacks, compromising - to varying degrees - at least 27 companies, including a Fortune 500 hospitality firm, a major US airline, an industrial supplies distributor and an online fashion retailer, with credit card records among the stolen data; that operator combined Cairn with two other open-source AI harnesses. Google has separately reported Cairn deployed in an attack against a US aerospace organization. The framework is being adopted in espionage and criminal operations alike.

Amazon's own conclusion is about tempo. An agent that finds and exploits a zero-day in minutes finishes before many security operations centers can manually triage an alert; the agents that failed against Redis ran continuously for 16 hours. The recommended posture is high-confidence detections wired to automated response, with humans overseeing outcomes rather than reviewing every action. Amazon shared two IP addresses from the campaign's records: 142.4.98.145 and 142.4.107.34.

The takeaway

The honest reading of this report is neither "AI has mastered hacking" nor "AI hacking is hype." The agents failed against a hardened database and a well-locked SSH front door, and needed an exposed, default-configured, authentication-free application to win. But "find the soft target and move faster than the defenders" is not a new strategy - it is the oldest one in the book, now running at machine speed and machine stamina, at near-zero marginal cost per attempt. The window between a vulnerable service appearing on the internet and an attacker knocking on it has been shrinking for years. This week, Amazon measured one such window. It was eleven minutes.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.