AI agents found and exploited a zero-day vulnerability
Amazon reports observing autonomous AI agents conducting an intrusion against university servers by finding and exploiting a zero-day vulnerability. The incident illustrates an escalation in attack sophistication, as well as the limits of human oversight that sometimes comes too late.
AI agents linked to a Chinese actor have, according to Amazon, discovered and exploited a zero-day vulnerability in MONAI Label, an open-source medical imaging application, to take control of a university server. The attack, observed over a 48-hour window starting September 8, demonstrates that tools designed to automate offensive tasks can now chain reconnaissance, exploitation, and escalation without constant supervision. Amazon alerted the project maintainers, who released a patch and a security advisory.
Des attaques qui apprennent en route, et ce n’est pas rassurant
According to Amazon's CISO, CJ Moses, the agents first attempted several unsuccessful intrusions before finding the right path. For nearly a day, they carried out multiple actions on university targets in East Asia: attempts on a Redis server, Log4Shell injection, SSRF attack against a co-located Qdrant instance, then various credential sprays against PostgreSQL, SSH, and other services. Another operation targeted a university GPU server for over six hours, using credentials built from 80 real names gathered from public sites, academic papers, and other documents. "Neither operation succeeded, yet together they show that the agents kept adapting for nearly a full day without human direction," Moses wrote.
The move to the decisive exploit was swift. The agents identified the application framework, mapped its API endpoints, then found a file upload endpoint with a filename parameter vulnerable to path traversal. About 11 minutes were enough to go from reconnaissance to arbitrary file write on the server. Later, the same vulnerability was used to launch a Python reverse shell executed as root. In this case, autonomy not only sped up the attack: it reduced the human reaction window to almost nothing.
Cairn, l’outil open source au cœur de l’opération
The attack relied on Cairn, an open-source pentesting framework that has become popular in multi-stage malicious operations. On September 10, Amazon spotted a server publicly hosting an instance of this tool, which allowed tracking the campaign. A human operator then provided a proxy to connect the agent to a second university network using the same vulnerable code. The detail matters: the AI did not act alone in absolute terms, but it handled the bulk of the offensive chain, with humans in the background.
This finding aligns with a previous report from Gambit Security, which also attributed the use of Cairn to a Chinese operator in at least 105 attacks, resulting in "to varying degrees" compromise of at least 27 companies. Again, multiple targets were hit, including a major Fortune 500 hospitality company, a large American airline, a significant private industrial distributor in the US, and an online fashion retailer. The convergence of these cases suggests that autonomous agents are no longer used solely to test defenses: they are entering data theft and espionage campaigns.
Ce que change cette affaire pour les défenseurs
The issue goes beyond the MONAI Label case alone. When an AI finds a vulnerability, exploits it, then adapts its strategy after several failures, security teams are no longer monitoring just scripts. They must track dynamic chains of actions capable of continuously varying targets and methods. This requires finer detections on logs, stricter segmentation policies, and rapid reduction of the exposed surface, especially on open-source services accessible from the Internet. Comfort here comes at a cost. And it is paid quickly.
- 48 hours of campaign observed starting September 8.
- 11 minutes between API mapping and arbitrary file write.
- 148 sequential actions against a university Redis server.
- 80 real names collected to fuel a credential spray attack.
- 105 attacks attributed by Gambit Security to another campaign linked to Cairn.