WORLDTECH NEWS Global technology intelligence.Contact
Cybersecurity INDEPENDENT REPORTS

Amazon says Chinese hackers used Cairn AI tool to discover apparent zero-day during attack on university

Close-up view of surveillance cameras during a snowfall in urban Москва setting.
Illustrative photo.Photo by Спиридон Варфаламеев on PexelsAmazon logo shown for identification only; no affiliation with or endorsement of WORLDTECH is implied.

What happened

CJ Moses, CISO of Amazon Integrated Security, said their investigation began on September 10, when they discovered a publicly accessible server with a publicly hosted instance of Cairn — an open-source autonomous AI penetration-testing framework. In this incident, the attacker provided IP addresses and asked Cairn to “obtain shell access” — meaning find a way to gain enough access to a system for various actions to be taken.

The only successful attempt occurred when the agents discovered a vulnerability in an open-source medical imaging AI application and exploited it. A China-linked threat actor (the person or group behind an attack) used the open source AI penetration testing tool Cairn to persistently attack two universities, and in one case, possibly discover a zero-day (a flaw already being used in attacks before a fix exists) bug, researchers said.

Threat actors used the open-source AI penetration-testing framework Cairn to find a previously unknown vulnerability in a medical imaging application. They were able to view the operational history of the platform, including previous prompts and more.

The operator wrote their prompts in Simplified Chinese and the timestamps were configured to China Standard Time. The targeted victim organizations were universities in East Asia.

Moses explained that Cairn is ideal for attackers because, when given a starting point and objective in plain language, it can coordinate multiple AI agents (AI that carries out multi-step tasks rather than answering one question) to complete a task. “Based on those inputs, the AI agents planned, executed, and adjusted their approach as they learned, without the operator choosing each next step,” Moses said.

One of the attempts involved 148 actions taken over 16 hours that were ultimately unsuccessful, including a failed attempt at exploiting the Log4Shell vulnerability.

Sources & evidence