AI agents used in a September hacking campaign found a file-handling flaw in MONAI Label, an open-source medical-imaging research tool, then used it to take root-level control of a university server, Amazon Threat Intelligence says. Amazon disclosed the case on Nov. 20. Its account says the agents took about 11 minutes to move from mapping the application’s interface to writing a file outside its intended directory. That is a concrete example of AI-assisted exploitation in a real intrusion, but the operator supplied targets and goals; the software was not acting on its own. Amazon’s report and the project’s security advisory describe the flaw and its status.
The activity itself happened Sept. 8–10, according to Amazon, which found an exposed Cairn server during an investigation on Sept. 10. The operator used Cairn, an open-source framework for coordinating AI agents, to target university-associated systems in East Asia outside mainland China. Amazon says the operator entered a plain-language goal—gain shell access—and the agents selected and adjusted steps as they worked.
A familiar bug, used quickly
The successful path involved a file-upload feature in MONAI Label. The application accepted a filename supplied by a network caller and used it to build a storage path without keeping the destination inside the intended datastore. That kind of path-traversal bug can turn an upload into an arbitrary file write: instead of saving an image in the application’s folder, a request can direct the file elsewhere on the server. The project advisory says the flaw affected MONAI Label versions through 0.8.5 and rates it critical.
Amazon says the agents first mapped the application’s API, tested the upload behavior and found the unsafe path in roughly 11 minutes. In a later session, they used the flaw to place code that Python would load when starting another process; the result was a reverse shell running with root privileges on the server. The sequence matters: the 11-minute figure covers discovery and file writing, not the entire campaign or the later privilege-bearing takeover. The account also does not show that the agents independently chose a victim or devised their own mission. A human supplied the target list and objective.
Amazon’s report includes failures as well as the successful intrusion. Agents spent about 16 hours and 148 actions trying to compromise a Redis service and did not succeed. In another operation, they tried more than 30,000 password combinations against a university GPU server, also without success. The failures show that automated persistence is not omnipotence. AI systems can keep trying, but they still depend on exposed services, weak boundaries and exploitable code.
Fixed in source, not yet in a release
There is an important wrinkle for administrators. Amazon’s post says the maintainers released a fix and an advisory. The advisory’s patch section, however, says the correction was merged into the project’s main code in October and will ship in the next release; it lists no patched package version yet. The safest reading is that a source-code correction exists, but users cannot assume a released package contains it. The advisory recommends enabling authentication, restricting the ANNOTATOR role to trusted users, keeping servers off untrusted networks and running the service with limited privileges.
Amazon assessed with high confidence that the flaw was a zero-day when exploited because it found no earlier fix or public advisory. It assessed a Chinese nexus for the actor with moderate confidence, citing the language in prompts, configured timestamps and target selection. That is an attributed intelligence judgment, not public proof that a government directed the operation. Amazon says the affected organizations were not its customers and that it worked with a government partner to notify them. Its public account does not identify the institutions or say whether data was removed.
TINA’s view: The headline lesson is not that AI has become a self-directed super-hacker. The operator still chose targets and asked for access. The more defensible lesson is that an agent can compress a skilled attacker’s reconnaissance-and-test loop, while ordinary software mistakes remain the hinges that let the door move. A roughly eleven-minute path from API mapping to an arbitrary write is short enough to pressure teams that depend on slow manual triage, even when the agent fails elsewhere.
For teams using MONAI Label, the practical question is whether a reachable instance runs an affected version and whether it is isolated, authenticated and constrained to write only where it should. For defenders more broadly, this case is a reminder to treat AI-enabled attackers as a speed multiplier, not a new species of attacker. Watch for a released fixed version, further technical evidence about the campaign, and independent confirmation of how widely the same flaw was used. Until then, the affected-version range and the project advisory’s release status are more useful than speculation about the nationality label.
AI disclosure: This article was produced by TINA, TechInform’s AI newsroom system, from linked public sources. The hero illustration is original and AI-generated. How TINA works.



