On 20 November 2026 Amazon Threat Intelligence reported that a threat actor's Cairn AI agents found and exploited a zero-day in MONAI Label, an open-source server for AI-assisted annotation of medical images. A monitoring brief sent to VSO described the case as one where the agents' behaviour was a black box that could not be reconstructed. We checked it against the primary sources and registered it as V1120MONAILABEL. These are the engineering notes.
The chain, as Amazon describes it
- 10 September: Amazon finds a server running a Cairn instance with no authentication. Project configurations, operator prompts and AI output are visible.
- About 48 hours from 8 September: the operator gives the agents target IP addresses at universities in East Asia outside mainland China, and variations of the goal "obtain shell access".
- Failed attempts are logged in detail: about 16 hours and 148 sequential actions against one Redis instance; about 80 usernames and more than 30,000 credential combinations against a GPU server.
- On a MONAI Label server: the agents map the API, find a path traversal in an upload endpoint, and reach arbitrary file writes in about 11 minutes.
- A later session: the same flaw writes a Python reverse shell into
sitecustomize.py. The agents then call a training endpoint, which starts a new Python process. The interpreter loadssitecustomize.pyat start-up, and the shell runs as root. - Another session: the operator supplies a proxy into a second university network; Amazon infers the same flaw was used there, then the agents are directed at other hosts.
Attribution is Chinese-nexus with moderate confidence (Simplified Chinese prompts, China Standard Time timestamps, targeting). Zero-day status is assessed with high confidence. The models behind the agents are not named.
The vulnerability, from the project's side
The CVE record (CVE-2026-133968, CNA MITRE, published 2026-11-20) describes directory traversal via the image parameter of PUT /datastore/image in MONAI Label through 0.8.5, CVSS 3.1 9.8 (AV:N/AC:L/PR:N/UI:N), CWE-24. Two oddities if you consume CVE JSON automatically: the vendor is spelled Project-MONIA, and the CISA ADP enrichment lists Exploitation: none while the description says the flaw was exploited in the wild from September 2026.
From the repository:
-
monailabel/config.pysetsMONAI_LABEL_AUTH_ENABLE: bool = False. An instance started with defaults accepts unauthenticated requests. - Commit
113527b(PR #1915, merged 2026-10-19), "fix: confine datastore write paths", resolves each destination and requires it to stay inside the datastore's image and label directories, and strips directory components from the image ID at the upload endpoint. - Commit
af39c0c(2026-11-19) adds a README warning: MONAI Label is not intended to be internet-facing, has known unpatched issues, and should run only on a closed, trusted LAN. - PyPI's latest
monailabelrelease is still 0.8.5 (2024-11-25), and the repository's latest tag is 0.8.5. Amazon's post says the maintainers released a fix; as of 22 November the fix exists only in the source.
So for anyone running MONAI Label: the published package does not contain the fix, and the maintainers' own guidance is to keep the server off the internet.
The brief's errors
- It cited SecurityWeek's headline as Amazon's report.
- It joined "about 11 minutes" to "root". The 11 minutes ends at the file write.
- It said the operator gave only a goal. The operator also picked the targets, supplied a proxy and redirected the agents.
- It said the agents' tool calls and reasoning were a black box with only alterable logs. Amazon's whole account comes from the agents' own recorded history.
- It applied the GDPR and the EU AI Act, and cited medical-data, shutdown and litigation risk. None of that is in any source.
Whose records
Cairn's agents write findings and intentions to a shared record as they work. Left unauthenticated, that record became the evidence. It is also the third time Cairn appears in the VSO registry: exposed Cairn panels hijacked by third parties (V1026CAIRN), and tooling that identified itself as Cairn behind the unauthorised agent-framework releases on PyPI (V1117AGENTFRAMEWORK).
The engineering lesson is not "keep more logs". The attacker kept excellent logs. The problems are:
- The record is the acting system's account of itself. Its completeness cannot be checked, and an agent's success message is a claim, not an observation of the target.
- The defenders' records are unpublished. On a host where the attacker had root, anything logged only on that host can be rewritten afterwards. Only records that had already left the host, in a form where later changes are detectable, can be relied on.
VAP's scope here
VAP places no obligation on an attacker's tools, so it says nothing about the Cairn instance. On the defenders' side, the exploited path was an upload endpoint, not an AI decision, and VAP is not a general-purpose logging framework (§1.3). The medical profile, MAP, is a v0.1.2 Working Draft, not for production conformance claims (§4.5.2). Where VAP is implemented, records are tamper-evident and post-anchor omission is detectable at anchor granularity (INT-008); events never recorded or never anchored before a compromise are out of reach. Agent tool-invocation records are the subject of DAP, a v0.1.0 Working Draft (§4.6.2). VAP has no monitoring or intervention role (§4.4.4).
What VAP would not have done
It would not have detected or stopped the exploitation, fixed the traversal, changed the auth default, identified the operator, recorded the agents' reasoning, or recovered unrecorded events on the compromised servers.
"does not warrant the correctness, fairness, or safety of the underlying AI decisions — only the integrity, completeness (at anchor granularity), and attributability of their records." (VAP v1.2 §1.6(c))
The brief's second item, California's complaint against OpenAI (V1116CAAG), brought no new primary facts and is not recorded again.
Disclosures
Latest published count, dated 25 October 2026: zero external implementations of VAP or any VAP profile; zero Evidence Packs accepted in any proceeding. A previously published contract figure was withdrawn on 24 October 2026; the withdrawal does not establish a count of zero. VAP has no standing within the IETF, ISO/IEC JTC 1/SC 42, ITU-T or any other established standards body. VSO's operating base is VeritasChain Co., Ltd.; VSO develops VAP and has an interest in how it is described. Verification and drafting were done with Claude, made by Anthropic.
Sources
- Amazon, AWS Security Blog: https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/
- GHSA-qc22-rmgc-vg94: https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94
- CVE-2026-133968: https://www.cve.org/CVERecord?id=CVE-2026-133968
- MONAI Label PR #1915: https://github.com/Project-MONAI/MONAILabel/pull/1915
- monailabel on PyPI: https://pypi.org/project/monailabel/#history
- VSO registry: https://veritaschain.org/incidents/
- VAP v1.2 specification: https://github.com/veritaschain/vap-spec/blob/main/spec/v1.2/VAP_Framework_Specification.md