// 1 CRITICAL · 2 ZERO-DAY · 4 CVE · 4 EXPLOIT IN THE LAST 24H→
Amazon Threat Intelligence documented Cairn framework AI agents autonomously discovering CVE-2026-133968 in MONAI Label and achieving root access on university servers during attacks in September 2026.

Amazon Threat Intelligence disclosed on November 18, 2026 the first documented case of open-source AI agents capable of autonomously discovering and exploiting a zero-day vulnerability in real-world software. The campaign, active from September 8 to 10, 2026 against universities in East Asia, led to the compromise of servers running MONAI Label, an open-source tool for AI-assisted medical image annotation.

The impact extends beyond the single exploit. The agents compressed the initial reconnaissance, discovery, and exploitation phase into roughly 11 minutes — faster than the response latency of many Security Operations Centers.

Key Takeaways
  • AI agents from the open-source Cairn framework autonomously discovered CVE-2026-133968 in MONAI Label, progressing from API mapping to arbitrary file write in approximately 11 minutes without step-by-step human intervention.
  • The root shell occurred in a subsequent session, not within the initial 11 minutes.
  • The vulnerability is a path traversal in the upload endpoint with a CVSS 9.8 (critical) score, exploitable without authentication in versions 0.8.5 and earlier.
  • Amazon attributes the attack to a China-nexus actor with moderate confidence, based on simplified Chinese prompts, CST timezone, and targeting of universities of strategic interest to China.
  • The fix was merged into source code on October 19, 2026, but no patched release was available at the time of disclosure.

How the Kill Chain Automation Works

The mechanism documented by Amazon breaks the traditional model of manual or semi-automated attack. According to The Hacker News reconstruction, the agents received a strategic objective from the operator — "obtain shell access" — and the choice of targets. From that point, the Cairn framework autonomously managed the entire initial sequence: mapping the MONAI Label API, identifying the vulnerable filename parameter in the upload endpoint, and achieving arbitrary file write outside the storage directory.

The vulnerability manifests as a path traversal. The server accepts a user-supplied filename and appends it to the storage path without validation, allowing injection of "../" sequences to traverse the filesystem. In a subsequent session, the agents planted a reverse shell in sitecustomize.py, which Python executes automatically at the start of every process. The trigger occurred via the training endpoint, which spawns a new Python process that loads the compromised file.

Persistence was achieved without sophisticated techniques: the nature of the injection into a Python runtime startup file guaranteed re-execution on every module load.

Speed as an Attack Vector

"An agent that discovers and exploits a zero-day in minutes completes its activity faster than many security operations centers can manually triage an alert" CJ Moses, CISO Amazon Integrated Security, via The Record

The 11-minute figure specifically covers the window from first API contact to arbitrary disk write, as SecurityWeek reports citing Amazon Threat Intelligence directly: "In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server." The root shell occurred in a subsequent session, not within this interval.

For SOCs still relying on manual alert triage, the intervention window on the initial phase is technically closed before an analyst can even open a ticket.

The Record adds a further element: in another session, the agents operated for 16 consecutive hours without interruption, attempting 148 actions against a Redis server. All failed, but the unlimited operation — without "breaks, sleep, or working-hour constraints" — introduces a structural asymmetry with human defenses.

Even failed operations are informative. The 30,000-plus credential attempts tested over more than 6 hours against 80 different usernames — all failed according to The Record — show that the agents do not stop at the first failure and that automation enables exploration of broad attack surfaces in reduced time.

Attribution and Its Limits

Amazon explicitly stated the limits of its analysis. Attribution to a China-nexus actor rests on behavioral and linguistic indicators — simplified Chinese prompts, timestamps in China Standard Time — combined with the nature of the targets, universities in East Asia outside mainland China, of known strategic interest to Beijing. Confidence is moderate. No APT group was named.

Amazon's documentation does not assert that the Chinese government directed the operation, nor that the attack falls within a known campaign. The dossier marks as unknown whether the second university targeted via proxy was actually compromised or only reached, and no public evidence emerges of clinical or patient data exfiltration.

Immediate Actions

The MONAI Label situation requires immediate and contextual actions:

  • Verify the presence of MONAI Label instances exposed on public networks, prioritizing version 0.8.5 or earlier, and isolate those not strictly necessary.
  • Evaluate the fix from source code available in the official repository, given that no patched release was available at Amazon's disclosure.
  • Enable authentication where possible, since MONAI Label disables it by default and the vulnerability is exploitable without an account.
  • Review upload policies to validate filenames and restrict writes to the intended storage directory.
  • Monitor Python startup files such as sitecustomize.py to detect reverse shell injections.

The gap between the fix merged on October 19, 2026 and the absence of an installable release represents a specific risk for environments that depend on distributed packages rather than manual builds from source code.

The Line Between Automation and Autonomy

The CVE-2026-133968 case does not represent a qualitative leap toward fully independent AI, but it documents a significant shift in the point of human intervention. The operator defined the goal and targets; the agents planned and executed the methods. As The Hacker News reported citing Amazon: "The agents worked with little step-by-step direction. The operator set the goal of getting shell access and chose the targets, and the agents selected the methods themselves."

Amazon assessed with high confidence that this was a zero-day, based on the absence of "any corresponding fix, issue, advisory, or prior mention in vulnerability databases." The discovery speed — 11 minutes from zero to file write — and the ability to operate for 16 consecutive hours introduce operational pressure on defenders that current SOC structures are not designed to absorb.

The Cairn framework is open-source and publicly available, making this attack model replicable by other actors. The combination of accessible tools, software targets with irregular maintenance, and a human response window measured in hours rather than minutes defines the perimeter of the problem for the coming months.

Sources


Sources and references
  1. securityweek.com
  2. thehackernews.com
  3. therecord.media
  4. bleepingcomputer.com
  5. startupfortune.com
  6. mlq.ai