Amazon Threat Intelligence reported on November 20, 2026 that a threat actor linked to China used an open-source autonomous AI framework to find and exploit a previously unknown vulnerability on its own, moving from mapping an application's API to writing files on the server in about eleven minutes. The case, documented in an AWS Security Blog post, is the clearest public example yet of AI agents compressing the timeline of an attack from days to minutes.
A Server Left Open
Amazon discovered the activity on September 10, 2026, when its threat intelligence team found a server hosting an instance of Cairn, an open-source autonomous AI penetration-testing framework, running without authentication. Because the operator left the instance publicly accessible, anyone on the internet could read its project configurations, the prompts the operator had written, and the output the agents produced. The prompts were written in Simplified Chinese, and timestamps in the executed scripts were set to China Standard Time.
Analysis of that instance showed a campaign aimed at IP addresses associated with universities in East Asia outside mainland China, spanning roughly forty-eight hours beginning September 8, 2026. Cairn was built as a research and problem-solving engine for defenders: give it a starting point and a plain-language objective, and it coordinates multiple AI agents to find a path to the goal. That same ease of use makes it attractive to attackers. The operator supplied a list of target addresses and variations of one instruction, "obtain shell access," then let the agents plan, execute, and adjust as they learned.
The Targets That Held
Two operations showed how long the agents would persist without human direction. Against a Redis instance at one university that allowed unauthenticated access, the agents worked for roughly sixteen hours across 148 sequential actions. They tried abusing Redis's replication feature to load unauthorized code, escaping its scripting sandbox, writing to its configuration files, and corrupting its memory. They also attempted a Log4Shell injection against another service on the same system. When those failed, they mapped the internal Docker network through a server-side request forgery flaw in a co-located Qdrant instance and sprayed credentials at PostgreSQL, SSH, and several other services. Nothing worked.
A separate operation ran for more than six hours against a university GPU server. The agents built a list of about eighty real usernames by researching faculty directories, lab websites, archived pages, and academic papers. They then tested more than thirty thousand credential combinations against the server's remote login and, on other systems across the campus network, an admin dashboard, a network storage device, a VPN server, and the campus portal. They never found a working login.
Eleven Minutes to a Root Shell
The outcome changed when the agents reached a target running an open-source medical imaging AI application. They identified the application's framework, mapped its API endpoints, and found an upload endpoint with a filename parameter. They tested that parameter for path traversal, and it was vulnerable. In about eleven minutes, they went from mapping the API to writing files anywhere on the server.
In a later session they used the same flaw to write a Python reverse shell into sitecustomize.py, a file Python imports automatically at interpreter startup. They triggered a training API endpoint to spawn a new Python process, the process imported the payload, and the agents got a shell running as root. At the time Amazon found it, the path traversal had no assigned security advisory and no patch. Amazon assesses with high confidence that it was a zero-day, because no fix, issue, advisory, or prior mention existed for the affected project.
The actor then stepped in. After the agents discovered the flaw, the operator handed them a proxy that tunneled into a second university network. The server at the other end ran the same vulnerable software, and Amazon infers the actor had used the zero-day to compromise it and establish the tunnel. The actor directed the agents through that tunnel to target other machines.
Attribution and the Fix
Amazon assesses with moderate confidence that the actor is a Chinese-nexus group, based on the Simplified Chinese prompts, the China Standard Time timestamps, and the choice of university targets. Amazon notified the project's maintainers, who released a fix and published a GitHub Security Advisory, GHSA-qc22-rmgc-vg94, on November 18, 2026. The affected organizations were not AWS customers, so AWS worked with a government partner to notify them. Separately, Gambit Security has documented a financially motivated group using Cairn against online retailers, which suggests the framework is now appearing in both espionage and criminal operations.
What It Means for Defenders
The lesson Amazon draws is about tempo. An agent that discovers and exploits a zero-day in minutes finishes its work faster than many security operations centers can manually triage an alert. The agents that failed against the hardened Redis target ran continuously for sixteen hours without breaks, sleep, or working-hour limits. Amazon's advice is to prioritize high-confidence detections that trigger automated response, with humans overseeing outcomes rather than reviewing every action.
The same tools are available to both sides. Cairn is open source, the model weights behind it are widely accessible, and the barrier to running an autonomous offensive agent is now a server and a prompt. For security teams, the practical response is to shorten the window between detection and action, because the attacker no longer needs a working day to find a way in.
If you are tracking how AI agents are changing security and business, more analysis like this is published at Crowe Endeavors.
Sources
"AI Agents Found Zero-Day In Minutes" (The System Prompt (YouTube), November 20, 2026) — https://www.youtube.com/watch?v=WhwH3QzR7Z8
"Chinese-nexus actor's Cairn AI agents found a zero-day on their own" (AWS Security Blog, November 20, 2026) — https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/
"Path Traversal in MONAI Label datastore upload endpoints (GHSA-qc22-rmgc-vg94)" (GitHub Security Advisory, November 18, 2026) — https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94
"Amazon says attackers' AI agents found and exploited a zero-day in 11 minutes" (VentureBeat, November 20, 2026) — https://venturebeat.com/security/amazon-says-attackers-ai-agents-found-and-exploited-a-zero-day-in-11-minutes-but-no-patched-release-is-available