Brief IA
›

Amazon: AI Agents Exploit MONAI Label Vulnerability

AI Tools·Tom Levy·

Amazon: AI Agents Exploit MONAI Label Vulnerability

Amazon: AI Agents Exploit MONAI Label Vulnerability
Key Takeaways
1According to Amazon, AI agents exploited an unknown vulnerability in MONAI Label on a university server.
2Approximately 11 minutes separated the API mapping from arbitrary file writing, according to CJ Moses.
3MONAI Label officials released a patch after being alerted by Amazon.
Why it matters : Amazon describes a discovery and subsequent autonomous exploitation of a vulnerability, as well as unsuccessful attempts and human intervention in another session.

According to CJ Moses, CISO of Amazon, Chinese spies used AI agents to discover and exploit a previously unknown vulnerability in MONAI Label, ultimately gaining complete control of a university server. The officials behind this AI application for medical imaging have since released a patch and a security advisory following Amazon's alert.

Agents Gain Administrator Rights

Approximately 11 minutes: this is the reported timeframe by CJ Moses between mapping the API and arbitrary file writing on the server. The agents recognized the application framework, identified its access points, and then examined the one intended for uploads. A test on the parameter designating the file revealed a path traversal, the vulnerability through which they accessed the machine.

According to the Amazon executive, this vulnerability was then used to launch a Python reverse shell with root privileges. During another session, once the flaw was discovered, the human operator provided the agents with a proxy establishing a tunnel to another university network. They thus connected to a server using the same vulnerable code.

Previous Attempts Had All Failed

Before this intrusion, the agents had spent about 16 hours on a Redis server belonging to another university, Moses reports. Their 148 successive actions included hijacking replication to introduce unauthorized code, attempts to escape the script sandbox, modifying configuration files, and memory corruption. None of these attempts were successful.

After these failures, they tried a Log4Shell injection against a different service and a server-side request forgery (SSRF) against a vulnerability in a Qdrant instance hosted in the same location. They also practiced password spraying against PostgreSQL, SSH, and other tools.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Meanwhile, a separate operation targeted a university GPU server for over 6 hours using this last technique. It relied on 80 real usernames collected from the institution's public sites, academic publications, and other documents. Moses indicates that these operations remained unsuccessful, while describing agents capable of adapting their attempts for nearly a day without human guidance.

Amazon Observes Cairn on a Public Server

Amazon's threat intelligence team spotted a publicly accessible instance of Cairn on a server on September 10, according to Moses. This discovery gave them visibility into a 48-hour campaign that began on September 8, targeting academic institutions in East Asia outside of mainland China. The instructions were written in simplified Chinese, and the scripts bore timestamps set to China Standard Time.

The tool used, Cairn, is an open-source penetration testing framework. It can chain the steps of an intrusion without human assistance. According to the Amazon executive, its use is increasing among cybercriminals.

Gambit Security Reports at Least 105 Attacks

Gambit Security had already reported another campaign conducted in September by a Chinese operator using Cairn. According to its threat intelligence team, this operator is responsible for at least 105 attacks and has compromised at least 27 companies to varying degrees. To steal credit card data and other sensitive information, this operator used Cairn and two other open-source AI tools.

Identified victims included a Fortune 500 hotel company, a major U.S. airline, a large private distributor of industrial supplies in the U.S., and an American online fashion retailer, among other businesses. For Moses, these cases indicate that Cairn is being used for both espionage and criminal activities.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.