Tag: Cairn

  • AWS: Cairn AI Agents Found a Zero-Day in MONAI Label (2026)

    AWS: Cairn AI Agents Found a Zero-Day in MONAI Label (2026)

    Takeaway: On November 20, 2026, Amazon Threat Intelligence published Chinese-nexus actor’s Cairn AI agents found a zero-day on their own. The post says a threat actor used the open-source Cairn framework to run autonomous hacking operations with little hands-on direction. After being given a goal of obtaining shell access, the agents found a previously unpatched flaw in an open-source medical imaging AI application and used it to gain full control of a server. Amazon says that, at the time of discovery, the issue had no matching advisory or fix in public vulnerability databases.

    A public GitHub security advisory for that application, GHSA-qc22-rmgc-vg94, identifies the software as MONAI Label and rates the issue Critical. It says versions through 0.8.5 are affected, lists no patched release, and states that a fix merged in the project will be included in the next MONAI Label release. The GitHub Advisory Database and NVD also list CVE-2026-133968 for this issue, with a CVSS 3.1 score of 9.8. The project advisory itself did not show a CVE identifier in the record we reviewed. This article stays with those official accounts. It does not reconstruct attack steps or publish exploit details.

    What AWS says it found

    According to the AWS Security Blog post, Amazon Threat Intelligence identified a publicly hosted Cairn instance on September 10, 2026. Cairn is described as an open-source autonomous AI penetration-testing framework. The instance was running without authentication, so project configurations, operator prompts, and model output were visible. Amazon says the operator wrote prompts in Simplified Chinese and that executed scripts used China Standard Time.

    From that instance, Amazon reconstructed a campaign that began on September 8, 2026, and ran for about 48 hours against IP addresses associated with universities in East Asia outside mainland China. The actor supplied target addresses and variations of a simple objective: obtain shell access. The agents then planned, ran, and adjusted their own next steps.

    Amazon assesses with moderate confidence that this was a Chinese-nexus threat actor, based on the language of the prompts, the timezone on the scripts, and the choice of university targets that Amazon says are likely of strategic interest to China. That is Amazon’s attribution, not an independent finding in this article.

    Not every operation succeeded

    The post is careful to show limits as well as speed. In one long-running operation, agents spent roughly 16 hours and 148 sequential actions against a university Redis instance that allowed unauthenticated access, plus other services on the same system. Amazon says none of those attempts succeeded. In another operation lasting more than six hours, agents gathered a large set of real usernames from public university material and tried a very large number of login combinations against remote access and other campus services. They did not find a working login.

    Amazon’s point is not that every target fell. It is that the agents kept adapting for many hours without a human choosing each step, including outside ordinary working hours.

    The MONAI Label issue

    The outcome was different against a target running the medical imaging application. Amazon says the agents identified the application, mapped its API, and in about 11 minutes went from that mapping to writing files on the server. In a later session, they used the same issue to obtain a root shell. Amazon assesses with high confidence that this was a zero-day at the time, because it found no corresponding fix, issue, advisory, or prior mention in vulnerability databases for the affected project.

    Amazon later observed the actor provide a proxy into a second university network that was running the same vulnerable software, and then direct the agents to use that path against other machines. AWS says it notified the project’s maintainers, who published a fix and a GitHub security advisory. The affected organizations were not AWS customers, and AWS says it had no direct path to notify them; it worked with a government partner to pass along the activity.

    The public MONAI Label advisory, published on November 18, 2026, describes a path-traversal weakness in datastore upload handling that can allow an arbitrary file write. The maintainers say that can lead to remote code execution or denial of service, depending on what the server process can write. Authentication is disabled by default, and the advisory says that, with authentication enabled, the Annotator role is sufficient to reach the affected functionality. Affected versions are listed as 0.8.5 and earlier. The latest published release we checked remains 0.8.5, and the advisory’s patched-version field is empty. The project says the correction is in pull request 1915 and will ship in the next release.

    Official workarounds until a release ships

    Until operators can upgrade to a release that includes the fix, the MONAI Label advisory recommends:

    • Enable authentication (MONAI_LABEL_AUTH_ENABLE=True) and do not grant the Annotator role to untrusted users.
    • Do not expose the MONAI Label server to untrusted networks.
    • Run the server as an unprivileged user whose write access is limited to the datastore directory.

    This article does not add configuration or mitigation steps beyond that official list.

    What AWS wants defenders to take from this

    CJ Moses, CISO of Amazon Integrated Security, frames the report as evidence that agentic offensive operations are already happening. An agent that finds and uses a previously unknown issue in minutes can finish its work faster than many security teams can manually triage an alert. Failed operations still show persistence: the Redis attempt ran for 16 hours without breaks.

    Amazon says organizations should prioritize high-confidence detections that trigger automated response, with people overseeing outcomes rather than reviewing every individual action. The post also notes that AWS active defense systems, including honeypots and network traffic analysis, are used to identify scanning and help restrict it on the AWS network at no extra cost to customers. Those controls apply to AWS’s own network; they are not a patch for MONAI Label.

    The post cites earlier research from Gambit Security, which Amazon says documented a separate, financially motivated actor using Cairn against online retailers. Together, Amazon argues, the cases show the same framework appearing in both espionage-style and criminal operations. This article does not expand that second case beyond Amazon’s summary.

    The AWS post also publishes indicators of compromise dated September 8 to 10, 2026. See the official blog for those addresses. For a different, earlier look at how prompts and secrets can leak through the AI supply chain — not this Cairn campaign — see our coverage of the Dragos canary-token study.

    What this article does not claim

    Amazon did not name the medical imaging project in the blog post; the public project advisory is what identifies MONAI Label. This article does not claim that a numbered MONAI Label release already contains the fix, that a CVE appears on the project advisory itself, or that every Cairn operation in the campaign succeeded. It also does not treat Amazon’s moderate-confidence attribution as proven, and it does not describe how to reproduce the vulnerability.

    Key facts

    ItemWhat official sources say
    AnnouncementAWS Security Blog post by CJ Moses, dated November 20, 2026.
    FrameworkCairn, an open-source autonomous AI penetration-testing framework, operated without authentication on a public server.
    WhenAmazon identified the server on September 10, 2026. The reconstructed campaign began September 8 and ran about 48 hours.
    TargetsIP addresses associated with universities in East Asia outside mainland China.
    Successful caseAgents found a then-unpatched issue in an open-source medical imaging AI application and used it to gain full control of a server, in about 11 minutes from API mapping to writing files.
    Affected softwareMONAI Label, per GHSA-qc22-rmgc-vg94. Versions through 0.8.5. No patched release listed.
    CVECVE-2026-133968 is listed in the GitHub Advisory Database and NVD (CVSS 3.1 9.8). The project advisory record we reviewed did not show a CVE ID.
    Fix statusMaintainers merged pull request 1915 and say the fix will be in the next release. Latest published release remains 0.8.5.
    AttributionAmazon, moderate confidence: Chinese-nexus actor, based on language, timezone, and targeting.

    FAQ

    What did Amazon announce on November 20, 2026?

    Amazon Threat Intelligence said it investigated a Chinese-nexus threat actor using Cairn AI agents to run hacking operations with little human direction, including the discovery and use of a previously unpatched flaw that gave full control of a server.

    What is Cairn in this report?

    Amazon describes Cairn as an open-source autonomous AI penetration-testing framework. An operator gives it a starting point and a goal in plain language, and it coordinates multiple agents to pursue that goal. The instance Amazon found was publicly accessible without authentication.

    Which software contained the zero-day?

    Amazon’s blog refers to an open-source medical imaging AI application and does not name the project. The public advisory GHSA-qc22-rmgc-vg94 identifies MONAI Label. Amazon says it reported the issue to the maintainers.

    Is there a patched MONAI Label release?

    The official advisory lists no patched versions and says the fix will be included in the next MONAI Label release. The latest published release we checked is still 0.8.5, which the advisory marks as affected. Until a fixed release is available, the maintainers publish workarounds around authentication, network exposure, and process privileges.

    Is this tracked as a CVE?

    Yes, in public vulnerability databases. The GitHub Advisory Database and NVD list CVE-2026-133968 and a CVSS 3.1 score of 9.8. The project’s own advisory, GHSA-qc22-rmgc-vg94, calls the issue Critical and did not list a CVE ID in the record we reviewed.

    Did every operation in the campaign work?

    No. Amazon describes lengthy failed attempts against a hardened Redis host and against university login services, alongside the successful case against the medical imaging application.

    Who does Amazon say was behind the campaign?

    Amazon assesses with moderate confidence that it was a Chinese-nexus threat actor, citing Simplified Chinese prompts, China Standard Time on executed scripts, and the university targets. This article reports that assessment and does not treat it as independently verified.

    What should MONAI Label operators do now?

    Follow the official advisory: enable authentication, withhold the Annotator role from untrusted users, keep the server off untrusted networks, and run it as an unprivileged user with write access limited to the datastore directory. Watch for a MONAI Label release that includes the fix in pull request 1915, and apply it when it ships. See the AWS Security Blog if you need the indicators Amazon published for September 8–10, 2026.