It took the AI agents nearly a full day of failing. They spent about 16 hours and 148 sequential actions trying to break into a university Redis server. They tried a Log4Shell injection, a server-side request forgery attack, and credential sprays against PostgreSQL and SSH. None of it worked. Then they found a flaw nobody had reported before, in an open-source medical imaging app called MONAI Label, and went from mapping the app’s API to writing files anywhere on the server in about 11 minutes. A Python reverse shell running as root followed. That, Amazon says, is the first documented case of AI agents autonomously discovering a zero-day vulnerability. The claim comes from a report by Amazon chief information security officer CJ Moses, shared with The Register on 20 November 2026.
The campaign ran for 48 hours from 8 September 2026 against universities in East Asia, outside Mainland China, The Register reports. Amazon’s Threat Intelligence Team found the operation on 10 September, when it spotted a publicly hosted instance of Cairn, an open-source AI penetration-testing framework, that the attackers had left accessible without authentication, giving Amazon visibility into what the agents were doing. The attackers had simply instructed the agents to “obtain shell access”, BleepingComputer reports, leaving them to discover and exploit the flaw on their own. The prompts the agents ran were written in Simplified Chinese, and the timestamps in their scripts were configured to China Standard Time. The Register describes the operators as suspected Chinese spies. Amazon attributed the campaign to a Chinese-nexus threat actor with moderate confidence, based on the language of the prompts, the activity timestamps, and the choice of targets, BleepingComputer reports. No group has been named, and no government or law-enforcement body has attributed the campaign.
The 48 hours before the breakthrough
The agents’ first target was a university Redis server, and they worked on it for about 16 hours. They tried to abuse Redis’s replication feature to load unauthorized code, escape its scripting sandbox, write to its configuration files, and corrupt its memory, running 148 sequential actions. None of these succeeded. When the Redis attempts failed, the agents tried a Log4Shell injection against a different service, a server-side request forgery attack against a vulnerability in a co-located Qdrant instance, and several credential sprays against PostgreSQL, SSH, and other tools. None succeeded either.
A separate operation ran for more than six hours against a university GPU server, compiling a list of 80 real user names found across the school’s public websites, academic papers, and other documents, and spraying credentials against it. BleepingComputer adds that the agents tried over 30,000 credential combinations in those six hours. It did not succeed. “Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction,” Moses wrote, as quoted by The Register.
That persistence is the part worth sitting with. The agents did not stop after the first failure, or the tenth, or the hundredth. They kept trying new approaches for nearly a full day, with no human telling them what to try next. The two failed operations added up to more than 22 hours of agent effort. The successful phase lasted 11 minutes.
Eleven minutes from API map to root shell
The breakthrough came against MONAI Label, an open-source medical imaging AI application. “The agents identified the application’s framework, mapped its API endpoints, and found an upload endpoint with a filename parameter,” Moses wrote. “They tested that parameter for path traversal, and the parameter was vulnerable. In about 11 minutes, the agents went from mapping the application’s API to arbitrarily writing files on the server.”
The flaw is now documented as GitHub Security Advisory GHSA-qc22-rmgc-vg94, “Path traversal in datastore upload endpoints allows arbitrary file write”. The PUT /datastore/image endpoint accepts a caller-controlled image query parameter used as the image ID without validation. The server joins it onto the datastore image directory and writes the uploaded file there, so a value such as ../../../tmp/x writes uploaded content to any location the MONAI Label process can write to. The PUT /datastore/label endpoint builds its destination the same way. The advisory says an arbitrary file write can lead to remote code execution, for example by overwriting application code or adding entries to .ssh/authorized_keys or cron, or to denial of service by overwriting configuration files.
Later, the agents abused the same flaw to execute a Python reverse shell running as root. BleepingComputer reports the agents wrote the reverse shell to the sitecustomize.py file, which loads automatically when Python starts, then triggered an API endpoint to establish root access. Amazon shared two IP addresses associated with the campaign, observed between 8 and 10 September: 142.4.98.145 and 142.4.107.34, BleepingComputer reports. In another session, after the agents discovered the zero-day, the human operator gave them a proxy that tunnelled into a second university network, and the agents connected to a server running the same vulnerable code. The Register does not say the second server was taken over, so this article does not claim two universities were breached. No university is named in either source, and neither source says what data, if any, was accessed or stolen from the compromised server.
Amazon notified the MONAI Label project’s maintainers, and they released a fix and a security advisory. The flaw was originally reported by @aws-sec-vdp, Amazon’s vulnerability disclosure handle. It is fixed in pull request #1915, and the fix will be included in the next MONAI Label release. MONAI Label’s maintainers released a security update addressing the flaw on 18 November, BleepingComputer reports. No CVE number has been assigned yet, and no release version or date has been given, so this article cites the advisory number, GHSA-qc22-rmgc-vg94, and not a CVE.
The door was open by default
The agents’ hard problem was finding the endpoint, not beating a login. MONAI Label ships with authentication disabled by default (MONAI_LABEL_AUTH_ENABLE=False), so in the default configuration any network client that can reach the server can exploit this flaw. With authentication enabled, the ANNOTATOR role is sufficient to reach the vulnerable endpoints. The fix exists, in pull request #1915, but it ships in the next release, and no version or date has been given. Internet-reachable deployments running the default configuration stay exploitable until that release lands. The patch advice writes itself: enable authentication, and do not expose the server to the internet.
Why 11 minutes is the real story
The 11-minute figure is the story, and the failed hours are what give it weight. Across the two failed operations, the agents put in more than 22 hours of work that went nowhere, about 1,320 minutes of effort, and then went from mapping an API to arbitrary file writes in about 11 minutes. That is roughly 120 times faster, as a comparison of effort. The zero-day did not just open a door. It collapsed the effort required by two orders of magnitude. Two caveats: the two phases hit different targets, and the failed operations ran in parallel, so this is a contrast of effort, not a single timeline and not a controlled speed test. Even so, the shape of it is what Moses is flagging. The agents do not get tired, and then they move at machine speed the moment they find a real flaw.
The Redis phase shows the other half of the capability. The agents ran 148 sequential actions in about 16 hours, one action every six and a half minutes, sustained without human direction for nearly a full day. The failed phase shows persistence. The 11-minute phase shows speed. The combination is the point: an agent that keeps adapting on its own, and then exploits a genuine flaw the instant it finds one.
One framework, two motives
Amazon’s discovery follows an earlier report by Gambit Security’s threat intelligence team, which also documented a Chinese operator using Cairn to carry out at least 105 attacks and compromise, “to varying degrees,” at least 27 companies. That campaign also occurred in September 2026. In the Gambit campaign, the attacker used two other open-source AI harnesses alongside Cairn to steal credit card records and other sensitive data from a Fortune 500 hospitality company, a major US airline, a large private US industrial supplies distributor, and a US online fashion retailer, among others. “Together, these cases show that this autonomous AI agent framework is being adopted in both espionage and criminal operations,” Moses wrote.
The numbers make the contrast concrete. The criminal operator ran at least 105 attacks and compromised at least 27 companies to steal card records. The espionage operator in the Amazon case ran one 48-hour campaign against East Asian universities and ended with a root shell through a zero-day the agents found themselves. Same open-source tool, opposite ends of the threat spectrum, and the espionage side is the one that found the zero-day.
The Amazon incident is the latest in a wave of disclosures about autonomous attacks. Gambit Security’s card-theft findings came in late September, NorForge tracked the retail e-commerce Cairn activity in October, Palo Alto Networks’ Unit 42 disclosed its CL-STA-1238 findings on 12 November, Amazon reported ARTEX agents stealing AI API keys on 13 November, Google’s Threat Intelligence Group reported its exploit_agent findings on 16 November, and the Amazon Cairn and MONAI Label report landed on 20 November. Six disclosures in under two months, from five different research teams, all describing the same shift: multi-stage intrusions running without humans. The Amazon campaign ran on 8 to 10 September, before almost every disclosure except Gambit’s. The campaign was already running while the industry was still documenting the earlier ones.
That wave needs one careful distinction. Google’s exploit_agent finding, reported on 16 November, was the first fully automated vulnerability discovery and exploitation Google had observed, but it worked by analysing patches to weaponise already-disclosed flaws. It did not autonomously discover a zero-day. The Amazon case is the first documented case in the sources read of agents autonomously finding a zero-day. That is the claim, with the Amazon attribution attached, not an absolute historical first.
Three other campaigns use the same framework and should not be confused with this one. Unit 42’s CL-STA-1238 campaign, disclosed on 12 November, used Cairn in 25 operations against at least 19 organisations and breached 10, including a US teachers’ union. TechReview covered it separately. NorForge’s October investigation picked up the retail thread, finding more than 20 exposed Cairn control panels. Its busiest cluster ran more than 1,100 attack projects against e-commerce and payment sites, trying to break into more than 1,200 sites and breaching more than 100 companies. And Amazon’s 13 November MadPot report covers ARTEX agents, not Cairn, stealing AI API keys from exposed servers. Same tool in some cases, different operators, different victims, different motives.
What to watch next
BleepingComputer confirmed the Amazon report on 20 November, so the story now has two major outlets behind it. Amazon’s own report has still not surfaced as a public URL. Whether a CVE number is assigned for GHSA-qc22-rmgc-vg94, and which MONAI Label release carries the pull request #1915 fix. Whether any university or MONAI project statement names the affected institutions or says what data, if anything, was taken from the compromised server. Whether the second university network, reached through the operator-supplied proxy, was also compromised. And whether the fix reaches the internet-exposed MONAI Label deployments still running with authentication disabled by default, since those stay exploitable until the next release lands.


