Chinese-state AI agents autonomously breach university server via zero-day

Suspected Chinese intelligence operatives deployed autonomous AI agents that identified and exploited a previously unknown vulnerability in an open-source medical imaging application, gaining full control of a university server in East Asia. The campaign spanning 48 hours beginning September 8 demonstrates how AI-assisted attack tools are lowering the barrier for sophisticated intrusion sets to operate without continuous human direction.

Amazon Chief Information Security Officer CJ Moses detailed the intrusion in a report shared with The Register. The attackers used Cairn, an open-source penetration-testing framework, to coordinate multi-stage attacks across multiple targets. Before discovering the zero-day, the agents spent approximately 16 hours and 148 sequential attempts attempting to compromise a university Redis server through replication abuse, sandbox escape, configuration writes and memory corruption — all unsuccessful. The same campaign also tried Log4Shell injection, a server-side request forgery against a Qdrant instance, and credential sprays against PostgreSQL, SSH and other tools without success.

In a separate operation lasting more than six hours, the agents targeted a university GPU server, compiling a list of 80 real usernames gathered from public websites, academic papers and other documents. None of those credential-spray attempts succeeded either. Moses noted that "agents kept adapting for nearly a full day without human direction" across these failed operations.

The breakthrough occurred when the AI agents identified a path traversal vulnerability in MONAI Label, an open-source medical imaging AI framework. The agents mapped the application's API endpoints, tested the upload endpoint's filename parameter for path traversal, and found it vulnerable. Within approximately 11 minutes of beginning API mapping, the agents could arbitrarily write files to the server. The same flaw was later abused to execute a Python reverse shell running as root.

In a follow-up session, a human operator provided the AI agents with a proxy tunnel into a second university network, connecting to a server running identical vulnerable code. The report indicates this pattern — autonomous discovery followed by manual augmentation — is becoming characteristic of campaigns attributed to suspected Chinese actors.

Gambit Security's threat intelligence team separately documented a Chinese operator using Cairn to carry out at least 105 attacks across 27 organizations including a Fortune 500 hospitality company, a major US airline, a large private US industrial supplies distributor and a US online fashion retailer. In that campaign, the attacker combined Cairn with two other open-source AI harnesses to exfiltrate credit card records and other sensitive data. Moses wrote that "together, these cases show that this autonomous AI agent framework is being adopted in both espionage and criminal operations."

The MONAI Label maintainers received coordinated disclosure from Amazon's Threat Intelligence Team and released a fix along with a security advisory (GHSA-qc22-rmgc-vg94). The vulnerability trackers can patch installations by updating to the latest framework release.

Autonomous probing before zero-day discovery

Before the AI agents located the path traversal flaw in MONAI Label, they invested considerable time in attacking other network services. The Redis server received 148 sequential actions over roughly 16 hours, including exploitation of the replication feature to load unauthorized code, attempts to escape the scripting sandbox, writing to configuration files and corrupting memory. None of these techniques succeeded. The agents also attempted a Log4Shell injection against a different service, a server-side request forgery attack against a vulnerability in a co-located Qdrant instance, and multiple credential-spray attempts against PostgreSQL, SSH and other tools. All failed.

A distinct six-hour operation targeted a university GPU server. The agents compiled a list of 80 real usernames drawn from public websites, academic papers and other documents, then launched credential-spray attacks against those accounts. None authenticated successfully. Despite the lack of results, these preliminary operations occupied substantial time and illustrate the patience attackers are willing to invest when operating without human oversight.

The report states that across these preliminary efforts, the agents maintained continuous operation for nearly a full day before achieving any foothold. This extended autonomous activity period is notable because it shows attack frameworks can sustain multi-vector probing without waiting for human instruction between stages.

MONAI Label path traversal flaw exploited

The MONAI Label application provides a web-based interface for annotating medical images using AI-assisted segmentation. The vulnerability exists in an upload endpoint that accepts a filename parameter without sufficient path validation. The AI agents discovered this endpoint, submitted crafted filename values that bypassed input checks, and wrote arbitrary files to locations outside the intended directory. This type of path traversal bug is particularly dangerous when the application later processes uploaded files with elevated privileges.

After gaining the ability to write arbitrary files, the agents uploaded a Python script that executed as root. This gave them full system control, including the ability to install persistent access mechanisms, exfiltrate data and use the compromised server as a pivot into other network segments. The entire process from initial API mapping to root-level shell execution took approximately 11 minutes.

The same vulnerability was subsequently used in a separate instance where a human operator provided the AI agents with a proxy tunnel leading into a second university network. The agents connected to a server running the same vulnerable MONAI Label code and repeated the path traversal and shell-execution sequence. This follow-on operation confirms that the initial compromise was not a one-off accident but reflects a repeatable pattern when the same vulnerable software runs on reachable hosts.

Broader campaign using Cairn and AI harnesses

Beyond the university server intrusion, Gambit Security's threat intelligence team documented a parallel campaign in the same month involving a Chinese operator who used Cairn to conduct at least 105 attacks across 27 organizations. Targets included a Fortune 500 hospitality company, a major US airline, a large private US industrial supplies distributor and a US online fashion retailer. In addition to Cairn, the attacker combined two other open-source AI harnesses to steal credit card records and other sensitive data.

Moses wrote that these cases collectively demonstrate the framework's dual use in espionage and criminal operations. The ability to autonomously cycle through reconnaissance, exploitation and post-exploitation actions without human pause changes the economics of attack campaigns. Defenders can no longer rely solely on known-exploit signatures because the same framework can discover and chain previously unknown vulnerabilities at speed.

What this means for developers

Open-source AI and machine learning frameworks have become target-rich environments. The MONAI Label case illustrates how quickly autonomous agents can progress from reconnaissance to code execution when an application exposes unsafe file-handling endpoints. Path traversal bugs in particular allow attackers to write beyond intended directories, and when combined with execution capabilities, can lead to full system compromise.

Developers maintaining machine learning infrastructure should review upload endpoints for path validation, run static analysis on file-handling code and consider rate-limiting or authentication on interfaces accessible to external users. The speed at which the agents — 11 minutes from API mapping to arbitrary file write — highlights the importance of assuming automated probing is already occurring.

Security teams should monitor for Cairn-like frameworks in network logs, as these tools generate distinctive multi-stage action sequences. The credential-spray patterns, sandbox-escape attempts and rapid API enumeration described in the report are detectable with proper telemetry. Unusual sequences of API calls, particularly those involving file uploads followed by script execution, should trigger investigation.

University and research computing environments face particular risk because of the open, collaborative nature of their infrastructure. Institutions running MONAI Label or similar medical imaging AI stacks should verify that installations apply the latest security updates and consider network segmentation to limit lateral movement if a single service is compromised.

The broader trend of AI agents operating across independent attack phases without human prompting signals that defensive strategies need to shift toward detecting autonomous behavior patterns, not just known exploit signatures. The 48-hour window from initial access to full server compromise demonstrates the pace at which these operations can unfold.