Hacking AI agents discover a critical vulnerability themselves

Amazon Threat Intelligence describes a campaign where autonomous AI agents conducted hacking operations with minimal human intervention. In one case, they discovered and exploited a zero-day vulnerability in open source medical software.

Amazon Threat Intelligence reported observing, in September 2026, a campaign conducted by an actor linked to China relying on autonomous AI agents to attempt shell access, i.e., remote control of a machine. The most striking point lies elsewhere: in one attack, these agents themselves found a zero-day vulnerability in an open source medical imaging application, then exploited it to gain root access. For defenders, the message is clear: the pace of attacks can now be measured in minutes, not on-call shifts.

Des attaques pilotées par objectifs, pas par clics

The technical framework used is called Cairn, an open source autonomous pentesting framework that coordinates multiple agents based on a natural language objective. In the instance observed by Amazon, the operator had left the server publicly accessible, which allowed examination of the history, prompts, and AI outputs. The requests were written in Simplified Chinese and the scripts bore timestamps in China time, two clues that led Amazon Threat Intelligence to attribute the activity with moderate confidence to a Chinese-origin actor.

The campaign targeted IP addresses associated with East Asian universities outside mainland China for about 48 hours, starting September 8, 2026. Cairn was originally designed to assist research and problem-solving on the defense side. In an attacker’s hands, the same logic becomes an exploration engine that tests, corrects, and relaunches tirelessly. No need for coffee, then, but the result is less friendly.

Une machine qui insiste longtemps sur une cible difficile

On a particularly hardened target, a university Redis server configured without authentication, the agents performed 148 actions over about 16 hours. Redis, an in-memory database often used to speed access to frequently consulted data, can become a valuable entry point into a network. The agents attempted to abuse replication, bypass the script sandbox, modify configuration files, and corrupt the service’s memory. They also tried to exploit Log4Shell on another service on the same machine.

When these paths failed, they changed angle. The co-located Qdrant instance served as a springboard to map the internal network via an SSRF vulnerability, an attack that forces a server to make requests to internal resources. This was followed by credential spraying against PostgreSQL, SSH, and other services. Nothing succeeded, but persistence counts as much as success: these agents continued working without human pause for almost a full day.

On a university GPU server, the method was even more massive. The agents first built a list of about 80 real user names by scouring the institution’s public sites, lab pages, web archives, and academic publications. Then they tested over 30,000 credential combinations against SSH, as well as an admin dashboard, network storage, a VPN server, and the campus portal. Again, no valid access was found.

La faille zero-day qui change la portée du dossier

The turning point came on an open source medical imaging application using an AI framework. The agents identified the API, spotted a file upload point with a name parameter, then tested this parameter for directory traversal. It worked. In about ten minutes, they went from API mapping to arbitrary file writing on the server.

In a later session, they reused the same flaw to drop a Python payload into sitecustomize.py, a file Python automatically loads at interpreter startup. They then triggered a training API to launch a new Python process, which imported the malicious code and opened a reverse shell executed as root. Amazon Threat Intelligence strongly believes this was a zero-day, due to the lack of patch, security advisory, or prior mention in the project’s vulnerability databases at the time of discovery.

Ce que l’attaque dit du marché de la cybersécurité

The rest shows how a vulnerability found by an agent can turn into a full operation. In another session, the actor provided a proxy allowing passage through a second university network; the server connected to this proxy ran the same vulnerable software, suggesting the flaw had already been used to establish the tunnel. The agents were then ordered to target other machines through this passage.

Amazon notified the project maintainers, who published a patch and a GitHub security advisory. The affected organizations were not AWS customers, requiring relay via a public partner to warn them. Gambit Security had already documented another use of Cairn by a profit-oriented group, this time against online retail brands. Two different uses, one trend: AI agent frameworks are moving from lab demos to real offensive operations.

Ce que les défenseurs doivent surveiller

The issue is no longer just attack capability, but loop speed. An agent that finds and exploits a zero-day in minutes can finish its work before a SOC has time to triage a first signal. On the other hand, failed attempts against Redis or SSH show that autonomy mainly brings endurance. For security teams, this pushes toward high-signal detections, automated responses, and human supervision focused on decisions, not every click. AWS’s active defensive measures also rely on global honeypots and network traffic analysis to identify scans and block some activity at the network level, at no extra cost to customers.

Key points

  • Cairn was used for autonomous offensive operations in September 2026.
  • Amazon reports an actor linked to China, moderate confidence.
  • A zero-day was exploited on a medical imaging application.
  • 148 actions in 16 hours on a university Redis server.
  • Over 30,000 combinations tested against a GPU server.
  • Gambit Security had already observed Cairn with a financial actor.

By the numbers

  • 48 hours — campaign window, starting September 8, 2026, according to Amazon Threat Intelligence.
  • 148 actions — sequence conducted against a Redis server over about 16 hours.
  • 30,000+ combinations — credential attempts against multiple university services.
  • 11 minutes — time from API analysis to arbitrary file writing.
  • 2 IP addresses — indicators of compromise published for logs from September 8 to 10, 2026.

To read