For the first time, Amazon Threat Intelligence observed autonomous AI agents independently finding a zero‑day vulnerability and using it to gain root access on a target system. The actors leveraged Cairn, an open‑source autonomous penetration‑testing framework, to run the entire attack chain with minimal human input, exposing a new risk vector for engineers and security teams.
What changed: autonomous agents can discover zero‑days
The Cairn instance was left publicly reachable without authentication, allowing anyone to view its configuration, prompts, and AI output. The threat actor supplied a simple objective—"obtain shell access"—and the agents autonomously mapped targets, identified a path‑traversal weakness in an open‑source medical imaging AI application, wrote a malicious sitecustomize.py payload, and triggered a training API endpoint to spawn a root reverse shell. The vulnerability had no public advisory or patch at the time, leading Amazon to label it a zero‑day.
Why it matters to AI, cloud, and security engineers
AI engineers building or integrating autonomous tooling must recognize that the same ease of use that benefits defenders can be weaponized. Cloud and platform engineers see that exposing AI services without authentication can leak operational intent and provide a foothold for adversaries. Security engineers need to account for attack patterns that do not follow traditional, human‑driven steps, such as rapid API enumeration, automated credential spraying, and multi‑step exploitation that spans containers, databases, and custom application endpoints.
Architectural and operational implications
- Public exposure of AI tooling: Unauthenticated access to frameworks like Cairn reveals prompts, configurations, and execution logs, which can be harvested for intelligence.
- Automated discovery of unknown flaws: Agents can iterate through API endpoints, test parameters for traversal, and validate exploitability without human guidance, shortening the time from discovery to compromise.
- Persistence attempts across diverse services: The campaign included 148 sequential actions against an unauthenticated Redis instance, SSRF exploitation of a co‑located Qdrant service, and large‑scale credential spraying against PostgreSQL, SSH, and other services. While those attempts failed, they demonstrate the breadth of techniques an autonomous system can employ.
- Supply‑chain exposure: The zero‑day existed in an open‑source medical imaging AI project. Organizations that adopt such components must track upstream security signals and consider additional hardening for file‑upload paths.
- Network tunneling via compromised hosts: After the agents discovered the flaw, the actor supplied a proxy tunnel into a second university network, indicating that compromised assets can be repurposed to extend reach without further manual effort.
Related CloudNinjas coverage: AWS.
What This Means For Practitioners
Evaluate any publicly exposed AI automation platform for authentication and audit logging; restrict access to trusted identities and monitor for unusual prompt patterns. Harden file‑upload endpoints by validating and sanitizing path parameters, and treat any open‑source component that processes external data as a potential attack surface. Implement detection for rapid, repetitive API calls and credential‑spray activity that may indicate autonomous agents at work. Finally, incorporate threat‑intel feeds that track AI‑driven campaigns to stay ahead of emerging tactics.

