Particle.news

AI Agents Exploit Zero-Day in MONAI Label to Root University Server

The campaign shows minimally supervised AI agents can map APIs and chain-exploit exposed services rapidly, prompting urgent calls for hardening and patching.

Overview

  • Amazon Threat Intelligence disclosed on Nov. 20–21 that autonomous AI agents run through the Cairn framework found a path-traversal zero-day in MONAI Label and used it to gain a root shell on at least one university server.
  • The flaw affected MONAI Label 0.8.5 and earlier when run with the project’s default no-login setting and let an unauthenticated client write files outside the intended storage folder.
  • Researchers reconstructed that the agents mapped the app and reached arbitrary file writes in about 11 minutes, and in a later session they planted a Python startup file that was loaded to produce a root-level reverse shell.
  • MONAI maintainers merged a source-code fix on Oct. 19 that forces uploads to resolve inside allowed folders, the issue was assigned CVE-2026-133968 with a 9.8 score, and reporting says a packaged patched release may not yet be widely available.
  • Amazon and others urge immediate steps: cut public exposure to MONAI Label, enable authentication, run the service under a limited account, check for unexpected Python startup files, and confirm any installed source includes the October fix because dual-use AI tools speed attackers and raise demand for automated defenses.