Amazon Threat Intelligence says a Chinese-nexus actor ran autonomous AI agents against university servers in East Asia for about 48 hours from September 8, 2026, and that during the run the agents found a zero-day in an open-source medical imaging AI application and used it to take full control of a server.

What Amazon found

On September 10, Amazon Threat Intelligence identified a server hosting a public instance of Cairn, an open-source autonomous AI penetration-testing framework, with no authentication in front of it. Because the operator left it open, the platform’s history was visible to anyone: project configurations, the actor’s prompts and the AI’s output. The prompts were in Simplified Chinese and the timestamps in the executed scripts were set to China Standard Time. Amazon assesses with moderate confidence that the actor is Chinese-nexus, citing those two details and the targeting of universities likely to be of strategic interest to China. That attribution is Amazon’s assessment, and the post does not name the actor or the universities.

The operator gave the agents a list of IP addresses tied to universities outside mainland China and a goal that varied only slightly from target to target: “obtain shell access.” From there, Amazon writes, the agents planned, executed and adjusted as they learned, without the operator picking each next step. The campaign ran for roughly 48 hours.

Advertisement

CyberTech Your brand belongs here. Reach the decision-makers who read CyberTech every day. Premium placements across the site and newsletter. Advertise with us

Two long runs that failed

The longest operation went at a Redis instance at one university that allowed access without authentication. Over about 16 hours and 148 sequential actions, the agents worked through a long list of known techniques for taking control of an open Redis server, and also tried 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 other services. None of it worked.

A second operation ran for more than six hours against a university GPU server. The agents built a list of about 80 real usernames from faculty directories, lab websites, archived web pages and academic papers, then tested more than 30,000 credential combinations against SSH and, on other systems across the university network, an admin dashboard, a storage device, a VPN server and the campus portal. They never found a working login. Amazon’s reading of the pair is that the agents kept adapting for nearly a full day without human direction, even though neither operation succeeded.

The run that worked

Against a target running an open-source medical imaging AI application, the result was different. The agents identified the application’s framework, mapped its API endpoints and found an upload endpoint whose filename parameter was open to path traversal. Amazon says about 11 minutes passed between the start of API mapping and arbitrary file writes on the server. In a later session the agents used the same flaw to get a reverse shell running as root.

When Amazon looked, the flaw matched no assigned advisory and was unpatched. Amazon assesses with high confidence that it was a zero-day, because it found no fix, issue, advisory or earlier mention in vulnerability databases for the affected project. Amazon notified the project’s maintainers, who released a fix and a GitHub Security Advisory. The affected organizations were not AWS customers, so AWS worked with a government partner to notify them.

In one further session the actor stepped in. After the zero-day was found, the actor handed the agents a proxy that tunneled into a second university network. The server behind it ran the same vulnerable software, and Amazon infers the actor had used the zero-day to compromise it. The actor then pointed the agents at other machines in that network.

Newsletter

Get the week's best tech coverage.

Free. Read by thousands of HR, tech, and business leaders.

Cairn has other users

Amazon notes that Gambit Security was among the first to document Cairn in a threat group’s hands. Its report on online retailers describes a separate, financially motivated operator, which CyberTech summarized in an earlier brief. Gambit reported that between September 10 and 15 alone, 105 attack projects were launched and at least 27 companies were compromised to varying degrees. In the report, written by Eyal Sela, Gambit’s Director of Threat Intelligence, the verdict on the tooling is blunt: “At very low cost, the AI tools demonstrated a level of patience, persistence, and creativity that most human attackers would be unlikely to sustain in this kind of attack, and achieved far greater results, far faster.”

Taken together, the two reports describe one framework in use by an espionage-linked actor and by a criminal one. Google reported a related shift in a framework that turns patches into exploits with no human in the loop.

What it means for the security leader

Amazon’s stated advice is about tempo. An agent that discovers and exploits a zero-day in minutes finishes before many security operations centers can triage an alert by hand, and the agents that failed against the Redis target ran for 16 hours without breaks, sleep or working-hour limits. Amazon says organizations should prioritize high-confidence detections that trigger automated response, with humans overseeing outcomes instead of reviewing each action.

In our reading, the failed operations carry practical detail too. The Redis target was configured without authentication, and the GPU server drew more than 30,000 credential attempts. Both are visible on a defender’s side before any exploit lands. A SOC that already alerts on high-volume failed logins across several services from one source has a signal that precedes the successful run. The success came through an application with a flaw that no advisory had yet described.

What to do

  • Inventory data stores and internal tools that accept connections without authentication, starting with anything holding research or student data, and close them first.
  • Alert on repeated failed logins against SSH, VPN, admin dashboards and portals from one source address, and decide in advance which of those alerts may block an address automatically.
  • Review each playbook and mark the high-confidence alert types where containment can run without an analyst clicking first.
  • Track maintainer advisories for open-source applications you expose to the internet, and check the project named in Amazon’s GitHub Security Advisory if you run a medical imaging AI application.
  • Search logs from September 8 to 10 for the two addresses Amazon lists: 142[.]4.98.145 and 142[.]4.107.34.

Source: Amazon Threat Intelligence, AWS Security Blog