Image Not FoundImage Not Found

  • Home
  • Cybersecurity
  • Amazon says autonomous Cairn agents found and exploited a MONAI Label zero-day in 11 minutes
A systems administrator checks terminal output on a laptop in a university server room, with equipment racks and indicator lights behind them.

Amazon says autonomous Cairn agents found and exploited a MONAI Label zero-day in 11 minutes

Amazon says a September intrusion campaign used autonomous AI agents to find and exploit a previously unknown flaw in MONAI Label, an open-source medical-imaging annotation application, reaching arbitrary file write in about 11 minutes after mapping the target API. That claim matters because it is not a benchmark or a vendor demo: it comes from a real incident targeting university research infrastructure, where speed and persistence, not just novelty, shaped the risk.

The question readers actually need answered is simpler than the hype cycle makes it sound: when an attacker can give an agent framework a target and a plain-language goal, what do defenders now have to automate, and which ordinary controls still stop the chain?

What happened in the campaign

In its incident write-up, Amazon Threat Intelligence says the campaign began around September 8 and ran for roughly 48 hours against IP addresses tied to universities in East Asia outside mainland China. Amazon discovered the operation on September 10 after finding a publicly reachable Cairn server with no authentication, exposing project configuration, prompts and agent output. The prompts were in Simplified Chinese, and script timestamps were set to China Standard Time, which is why Amazon describes the actor only as Chinese-nexus, with moderate confidence.

The human operator did not vanish from the loop. Amazon says the person supplied targets and the goal of getting shell access, then let multiple Cairn agents plan, test and adapt across the environment. Later, the operator also provided a proxy into a second university network and directed the agents toward additional hosts. That is a meaningful distinction. The story is not that AI independently chose a campaign and ran it end to end. It is that human intent was translated into a long-running machine-led cycle of reconnaissance, exploit testing, retrying and pivoting.

The most consequential step involved MONAI Label. According to Project MONAI’s advisory, versions 0.8.5 and earlier were affected by a path-traversal flaw in the PUT `/datastore/image` endpoint, with related construction in PUT `/datastore/label`. The service accepted a caller-controlled identifier and joined it to the datastore path without ensuring the resolved location stayed inside the intended directory. In practice, an attacker who could reach the service could write files anywhere the MONAI Label process had permission to write. Authentication was disabled by default, and when enabled, the ANNOTATOR role was sufficient.

Why the 11-minute claim matters — and what it does not mean

The eye-catching number is specific. Amazon says the agents went from application and API mapping to arbitrary file write in about 11 minutes. That is not the full intrusion timeline, not the time to root, and not a general benchmark for every agentic attack. Root shell came later, in a separate session, when the agents used the same flaw to place a Python reverse-shell payload in `sitecustomize.py` and then triggered a new Python process.

That distinction does not make the episode less serious. It makes it more useful. The compression point is not “AI instantly owns any server.” The compression point is the space between exposure and first working exploit attempt. In many environments, especially research networks and smaller labs, security operations are still built around human review of alerts, tickets and after-hours escalation. If an automated attacker can enumerate an API, identify a path traversal, test write locations and land arbitrary file write before anyone has triaged the first noisy signal, the old workflow is already behind.

Just as important, Amazon’s own account shows limits. The agents spent about 16 hours and 148 sequential actions trying and failing to compromise an unauthenticated Redis instance, including replication, configuration changes, memory-corruption attempts, Log4Shell and a co-located Qdrant SSRF path. In another failed operation, they spent more than six hours against a university GPU server, collecting about 80 real usernames from public university material and trying more than 30,000 credential combinations across SSH and other services. Those details matter because they show persistence and adaptation, not magical reliability. As The Register reported, the successful MONAI chain sat alongside long, unsuccessful efforts.

What defenders now need to automate

The first priority is inventory and exposure reduction. Universities, hospitals, AI labs and vendors running MONAI Label or similar ML services should identify internet-facing deployments, especially MONAI Label 0.8.5 and earlier, and verify whether authentication is actually enabled. Too many research tools are treated as “internal” because they were meant for a trusted workflow, even when they are reachable enough for an attacker to enumerate their APIs from outside.

The second priority is preventing a single application bug from becoming host takeover. In Amazon’s account, arbitrary file write became far more dangerous because the process could write to a Python startup-hook location and an API path could trigger a new Python process. Least-privilege service accounts, tighter filesystem permissions, network isolation and strict boundary validation on upload paths all work here. They are not glamorous controls, but they are exactly the kind that break an agent’s momentum. A persistent tool can keep trying. It cannot wish away a service account that lacks permission to modify startup hooks or a segment boundary that blocks lateral movement.

The third priority is automated detection and narrow containment around high-signal sequences. Defenders should alert on rapid API enumeration followed by path-traversal patterns, unexpected writes outside datastore directories, changes to `sitecustomize.py`, reverse-shell behavior, or long-running credential-spray activity crossing protocols. The key operational change is speed: some responses now need to happen while a human supervises, not only after a human decides. Rate-limiting a client, quarantining a workload, revoking a token, disabling an exposed upload route or isolating a research subnet are the kinds of bounded actions that can buy time without handing all judgment to automation.

There is a final, less discussed lesson in Amazon’s disclosure: agent frameworks themselves are now attack surface. The company found the campaign because the Cairn server was publicly reachable with no authentication, exposing project data and agent output. Organizations experimenting with open-source offensive or testing agents should treat their control panels and logs like production systems, not lab toys. Restrict access, limit egress, and log agent actions well enough to reconstruct what happened if the tool is misused or exposed.

The broader business impact is likely to arrive through governance as much as incident response. Smaller research institutions and open-source maintainers rarely have enterprise SOC budgets, yet they run attractive, specialized software on GPU and imaging infrastructure. That makes this less a story about one framework than about the economics of defense. Procurement reviews, cyber insurers and cloud-baseline programs may increasingly ask not just whether an organization patches, but whether it can detect, isolate and recover fast enough when a reachable flaw is probed continuously by software.

Amazon also published two IP addresses seen in records from September 8 to 10 — 142.4.98.145 and 142.4.107.34. They are useful as historical indicators, not as a complete strategy. The larger takeaway is that defenders should assume the next agentic campaign will not announce itself with a dramatic payload. It may begin with ordinary API mapping against an ordinary service that was never meant to be exposed, then move faster than a conventional ticket queue can.