Skip to content
AI Security Wire

Published

- 7 min read

By

Amazon: Cairn AI Agents Found a Zero-Day and Rooted a University Server

img of Amazon: Cairn AI Agents Found a Zero-Day and Rooted a University Server

Incident Classification: Partially Disclosed | Incident Type: Reported | Severity: High (our assessment: confirmed impact was limited to unnamed university servers, though the CVE scores Critical) | Sector: Education, medical imaging research software | Date of Activity: 8 to 10 September 2026 | Documented by Amazon Threat Intelligence: 20 November 2026

Amazon Threat Intelligence says a threat actor aimed the open-source Cairn agent framework at university networks in East Asia, supplying target IP addresses and variations of the goal “obtain shell access”, and that the agents found an apparent zero-day in a medical imaging AI application on their own. AWS says it saw this because the operator left the Cairn server open to the internet without authentication, exposing the prompts, project configurations and AI output. The flaw now has a CVE record and a project advisory, and the fix is on the project’s main branch, but the newest PyPI release we found, 0.8.5, is inside the affected range. Victims are unnamed and AWS is the only source for the campaign, which is why we classify it as partially disclosed.

Incident Summary

FieldDetail
ToolCairn, an open-source autonomous AI penetration-testing framework (per AWS)
Operator exposurePublicly reachable Cairn instance with no authentication, identified 10 September 2026
TargetsIP addresses tied to universities in East Asia outside mainland China, about 48 hours from 8 September
Initial accessPath traversal that AWS’s post links to MONAI Label advisory GHSA-qc22-rmgc-vg94 (CVE-2026-133968), leading to a root reverse shell
Failed operations16 hours and 148 actions against an open Redis instance; 6+ hours and 30,000+ credential combinations against a GPU server
AttributionChinese-nexus, moderate confidence (AWS); no group named
VictimsNot named; AWS says the organisations were not AWS customers and were notified through a government partner

Timeline

  • 8 September 2026: Campaign against the university IP addresses begins, running about 48 hours (AWS).
  • 10 September: Amazon Threat Intelligence identifies the exposed Cairn server and its history.
  • Date not stated by AWS: The flaw is reported privately by the AWS vulnerability disclosure account against the MONAI repository (project advisory).
  • 19 October: MONAI Label pull request 1915 is merged (GitHub).
  • 18 November: The project advisory is published (GitHub).
  • 20 November: AWS publishes its report. The CVE record CVE-2026-133968 is published the same day, and The Record covers the story.

What the Agents Did

AWS describes several operations that varied widely in complexity and outcome. Two failed. AWS says they show both the persistence and the capability of the agents.

The longest targeted a Redis instance that allowed unauthenticated access. Over roughly 16 hours and 148 sequential actions the agents worked through known takeover techniques: abusing replication to load unauthorized code, escaping the scripting sandbox, writing to configuration files and corrupting memory. They also tried Log4Shell against another service on the same host. When that failed they mapped the internal Docker network through a server-side request forgery bug in a co-located Qdrant instance and sprayed credentials at PostgreSQL, SSH and other services. Nothing worked.

A second operation ran for more than six hours against a university GPU server. The agents built a list of about 80 real usernames from faculty directories, lab websites, archived pages and academic papers, then tried more than 30,000 credential combinations against SSH and, on other systems, an admin dashboard, a storage device, a VPN server and a campus portal. No working login was found.

The Zero-Day

The third target was different. The agents identified the application’s framework, mapped its API and found an upload endpoint with a filename parameter. They tested it for path traversal and it was vulnerable. AWS says it took about 11 minutes to go from mapping the API to arbitrary file writes.

In a later session the agents wrote a Python reverse shell to sitecustomize.py, which Python imports at interpreter start. They then triggered a training API endpoint that spawned a new Python process, which loaded the payload and opened a reverse shell running as root.

AWS assesses 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 for the affected open-source project”. The Record calls it an apparent zero-day. AWS’s post does not name the product: it links the MONAI Label advisory as the project’s fix and describes a “medical imaging AI application”. The advisory text says it was originally reported by @aws-sec-vdp against the MONAI repository and re-reported against MONAI Label to generate the CVE; that this handle is AWS’s disclosure account is our inference. The match is well supported, but it is our reading. AWS says the vulnerable parameter was a filename parameter, while the advisory describes an image parameter. The advisory’s description is consistent with AWS’s account: PUT /datastore/image takes a caller-controlled image parameter, joins it onto the datastore directory and copies the uploaded file there without checking the resolved path. A value such as ../../../tmp/x writes anywhere the process can write. The advisory says the PUT /datastore/label endpoint builds its destination the same way, although the CVE record names only /datastore/image.

In another session the human actor stepped in. After the zero-day was found, they gave the agents a proxy into a second university network. AWS says the server behind the proxy ran the same vulnerable software and infers the actor had used the flaw to compromise it.

Severity and Status of CVE-2026-133968

  • Affected: MONAI Label through 0.8.5 (CVE record, MITRE as CNA).
  • Score: CVSS 3.1 base 9.8 (AV:N/AC:L/PR:N/UI:N), set by the CNA. NVD lists it as a Secondary score from [email protected] and, when we looked, showed no NVD-primary score. CWE is CWE-24 (Secondary).
  • Default exposure: the advisory says authentication is off by default, so any client that can reach the server can exploit it. With authentication on, the ANNOTATOR role is enough.
  • Fix: merged in pull request 1915 on 19 October 2026 and described as shipping in the next release. The newest PyPI release is 0.8.5, uploaded in November 2024 and inside the affected range, so as of 21 November 2026 we see no PyPI release outside it.
  • KEV: no evidence found as of 21 November 2026 in the CISA KEV feed (catalogVersion 2026.11.20, released 20 November at 17:48 UTC, before the CVE record was published at 18:31 UTC), so the check says little yet. CISA’s SSVC entry lists Exploitation as none, Automatable as yes and Technical Impact as total. The Exploitation value differs from the CVE record’s “exploited in the wild” wording.

Indicators

AWS lists two addresses that appear in records dated 8 to 10 September 2026: 142[.]4.98.145 and 142[.]4.107.34. AWS does not say what role they played, so treat them as context for hunting, not attribution.

Remediation and Response

From the project advisory:

  1. Enable authentication with MONAI_LABEL_AUTH_ENABLE=True and do not give the ANNOTATOR role to untrusted users.
  2. Do not expose the server to untrusted networks.
  3. Run it as an unprivileged user with write access limited to the datastore directory.

Our recommendations, not from the sources:

  1. Check each MONAI Label host for an unexpected sitecustomize.py or other recently modified files in Python site directories, and for Python child processes started by the training endpoint.
  2. Review egress from research servers for outbound connections you did not authorize.
  3. Treat any unauthenticated Redis, Qdrant or AI service in the same network as reachable by a patient, automated attacker, since AWS shows these agents kept trying for hours.

What This Means

AWS draws one operational point: an agent that finds and exploits a zero-day in minutes finishes before many security operations centers triage an alert, and the failed Redis agents ran 16 hours without a break. AWS advises prioritizing high-confidence detections that trigger automated response, with humans overseeing outcomes.

For earlier Cairn coverage on this site, see our pieces on Unit 42’s CL-STA-1238 report and the Reuters report. AWS says Gambit Security documented a separate, financially motivated actor using Cairn against online retailers. Whether these are one operator or several is not established in the sources we reviewed. Our assessment: the notable detail is not the tooling but the target. Research software that ships with authentication off is exactly what a patient automated scanner finds, though AWS does not say whether the victim servers ran with authentication enabled.

Limits: the account of what the agents decided rests on AWS’s reading of the operator’s logs, and the sole source for the campaign is AWS. The Record’s coverage repeats the AWS findings.

References

Frequently Asked Questions

Did the AI agents really find the vulnerability without a human choosing each step?
That is Amazon Threat Intelligence's account, based on the operator's own exposed Cairn history. The operator supplied IP addresses and a goal, and AWS says the agents planned, executed and adjusted without the operator choosing each next step. The same report says a human did step in later, handing the agents a proxy into a second university network. We have only AWS's reconstruction, not the raw logs.
Is CVE-2026-133968 exploited in the wild according to the authorities?
Sources disagree on wording. The CVE record text says the flaw was exploited in the wild starting in September 2026, and AWS assesses with high confidence that it was a zero-day. CISA's SSVC entry, timestamped 20 November 2026, lists Exploitation as none, and we found no KEV listing as of 21 November 2026. We cannot reconcile these, so treat the exploitation claim as AWS's finding.
What should teams running MONAI Label do now?
The project advisory says the fix will ship in the next release, and the newest PyPI release we saw was 0.8.5 from November 2024. Until a fixed release exists, the advisory's workarounds apply: enable authentication, restrict the ANNOTATOR role, keep the server off untrusted networks and run it as an unprivileged user limited to the datastore directory.
Which attribution does AWS give?
AWS assesses with moderate confidence that the actor is Chinese-nexus, citing Simplified Chinese prompts, China Standard Time timestamps in executed scripts and a target set of universities likely of strategic interest to China. It names no group and did not name the universities.