Amazon Threat Intelligence said on November 20 that a Chinese-nexus actor used AI agents to attack university systems with little human help. The agents ran on Cairn, an open-source AI penetration-testing tool. Amazon found the actor's server open to the internet on September 10. Its records showed a campaign of about 48 hours from September 8, aimed at universities in East Asia outside mainland China.
One target ran an open-source medical imaging AI app. Its upload feature took a filename without checking it, a flaw called path traversal. The agents found it by testing that parameter. It let them write files outside the intended folder. They planted code in sitecustomize.py, which Python loads at startup, then triggered a new Python process. It loaded the code and gave them root control. Amazon has high confidence the flaw was a zero-day, with no fix or public record. The maintainers have since shipped a fix. Amazon has moderate confidence the actor is Chinese-nexus.
Two other operations, against a Redis database and a GPU server, failed. The success came from a flaw with no public record. That suggests patch lists miss an unlisted bug in a niche open-source app. This is interpretation. Leaders can ask: which open-source AI apps run inside our network, who owns their upload endpoints, and what contains a high-confidence alert automatically?