Suspected China-linked attackers pointed autonomous AI agents at university networks, and the agents found and exploited a previously unknown software flaw to take full control of a server. That is the account of Amazon's chief information security officer, CJ Moses, who published the report this week

1

2

.

Source: The Register

Source: The Register

The agents were given one instruction: obtain shell access. They did the rest themselves, per Amazon Threat Intelligence

2

. Per Amazon, the attribution to suspected Chinese attackers carries moderate confidence, resting on the Simplified Chinese prompts, the timestamps and the choice of targets. The attackers ran a roughly 48-hour campaign beginning on September 8, 2026, hitting universities in East Asia outside Mainland China

1

2

.

A day of failures, then 11 minutes

The agents' early attempts failed. They spent about 16 hours and 148 actions trying to compromise another university's Redis server, abusing its replication feature to load code, escaping the scripting sandbox and corrupting memory. It did not work

1

2

. They also tried a Log4Shell injection, a server-side request forgery attack against a co-located Qdrant instance, and credential sprays against PostgreSQL and remote login services. In a separate six-hour operation, the agents compiled 80 real usernames from public sources for a credential spray against a university GPU server. BleepingComputer puts that spray at more than 30,000 credential combinations tried. Neither attempt got the agents in

1

2

. "Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction," Moses wrote

1

.

Then they hit MONAI Label, open-source software that helps researchers label medical images, and found a path traversal flaw nobody knew about. From mapping the app's API to writing arbitrary files took about 11 minutes, per Amazon's account of one session. By that account, the agents later used the flaw to run a Python reverse shell as root

1

2

.

Source: VentureBeat

Source: VentureBeat

Humans did intervene at points. In a later session the operator supplied a proxy into a second university network with the same vulnerable software, and pointed the agents at other machines

1

2

.

How the campaign surfaced

Amazon found the campaign on September 10 through the attackers' own mistake: a Cairn instance they had left running on the public internet without a password, exposing their operations

1

2

. Cairn is a legitimate open-source pentesting framework that coordinates multiple AI agents, built for research and defense, and it has become popular with criminals

1

2

.

Amazon reported the flaw to MONAI Label's maintainers, who put out a patch and security advisory on November 18. Organizations running MONAI Label need to update

1

2

.

Other operations have used Cairn too. Per Gambit Security, a separate intelligence firm, a Chinese operator used Cairn in at least 105 attacks in September that compromised, to varying degrees, at least 27 companies. Targets included a Fortune 500 hospitality company and a major US airline. Google has reported a Cairn attack against a US aerospace organization

1

2

. Amazon's warning is that agents find and exploit flaws faster than security teams can react, and it wants defenses automated to match

1

2

. "Together, these cases show that this autonomous AI agent framework is being adopted in both espionage and criminal operations," Moses wrote

1

.

Source: BleepingComputer

Source: BleepingComputer

© 2026 TheOutpost.AI All rights reserved