Amazon says Cairn AI agents found a zero-day, names no project; a possible match, MONAI Label, has no patched version
Amazon says Cairn AI agents found and used a zero-day against a medical imaging AI application from 8 September 2026, and does not name it. The MONAI Label advisory of 18 November, 71 days later (derived), lists no patched version; its CVE record says exploited, CISA-ADP says none.
By Parminder Kumar Sharma · LinkedIn · · 20 min read
71 days from the start of Amazon's campaign to a project advisory, and Amazon never names the project
Amazon Threat Intelligence says a Chinese-nexus actor's Cairn agents were pointed at addresses of universities in East Asia over an "approximately 48-hour period beginning September 8, 2026", and that in one of three operations the agents used a zero-day in "an open-source medical imaging AI application" (AWS Security Blog, 20 November 2026, 17:30 UTC). The MONAI Label project's own security advisory is dated 18 November 2026. That is 71 days after the campaign began (derived: 8 September to 18 November) and 2 days before Amazon's post. A CVE record, CVE-2026-133968, was published at 18:31 UTC on 20 November, 1 hour 1 minute 25 seconds after Amazon's post (derived).
None of that makes MONAI Label Amazon's application. Amazon does not print the name. The link in this article is an inference from a match of features, set out in a table below, and Amazon naming the project would settle it. The 71 days is also a gap from the start of the campaign, not from a measured first exploitation: Amazon places the zero-day inside a window of about 48 hours from 8 September, so if the exploitation fell inside that window the gap to the advisory is between 69 and 71 days (derived).
This site read the pages on 21 November 2026 between about 08:42 and 08:48 UTC. Several are living pages (a pull request, an advisory, a CVE record), so each is quoted as saved at that time. Amazon sells cloud and security services, and the post describes AWS's own defences; that interest is stated here once. The pages name no victim, and this article names none.
What Amazon says happened, operation by operation
Amazon says that on 10 September 2026 it identified a server with a publicly hosted instance of Cairn, "running without authentication", so that the platform's history, including project configurations, prompts and AI output, was visible to anyone on the internet. The operator wrote prompts in Simplified Chinese and the timestamps in executed scripts were set to China Standard Time. Of Cairn the post says it "was built as a research and problem-solving engine for defenders": given a start point and a goal in plain language, it coordinates several AI agents. In this campaign the goal was variations of "obtain shell access" against a list of target addresses.
Amazon describes three operations. Redis: roughly 16 hours and 148 sequential actions against a Redis instance that allowed access without authentication, then credential sprays against other services on the same network. None succeeded. GPU server: more than 6 hours against a university GPU server, with a list of about 80 usernames built from public pages, and more than 30,000 credential combinations tested against remote login and several other systems on the network. Amazon says they never found a working login. Medical imaging application: the agents identified the framework, mapped the API, found an upload endpoint with a filename parameter, tested it for path traversal and found it vulnerable. In about 11 minutes they went from mapping the API to arbitrary file writes. In a later session the post says the same flaw gave a reverse shell running as root. This site does not reprint the mechanism Amazon describes.
Three rough figures follow from Amazon's own numbers, each derived and each only as good as its "roughly", "about" and "more than". 148 actions in roughly 16 hours is about 9 an hour, or one every 6.5 minutes (148 divided by 16 is 9.25; 960 minutes divided by 148 is 6.5). More than 30,000 credential combinations in more than 6 hours is about 83 a minute (30,000 divided by 360), but the combinations were spread over several systems, so the rate against any one service was lower, and because both figures are lower bounds the true rate is not bounded either way. Spread evenly over about 80 usernames, 30,000 combinations is about 375 attempts per name; Amazon does not say they were spread evenly. For comparison, the successful step took about 11 minutes.
Amazon's reasoning for "zero-day" is an absence: it assesses with high confidence that the flaw was one "based on the absence of any corresponding fix, issue, advisory, or prior mention in vulnerability databases for the affected open-source project". It attributes the campaign with moderate confidence to a Chinese-nexus actor, from the Simplified Chinese prompts, the China Standard Time timestamps and the targeting of universities "likely of strategic interest to China". Both the confidence levels and the evidence are Amazon's; none of the other saved pages tests them.
Where Amazon says the agents acted alone, and where a person acted
The post's headline says the agents "found a zero-day on their own" and its opening says the actor used autonomous agents "to conduct hacking operations with minimal human involvement". In the body Amazon says the agents "planned, executed, and adjusted their approach as they learned, without the operator choosing each next step". The same post also records a person at the start and in the middle. The table sets them side by side.
Who did what, as Amazon's post of 20 November 2026 states it. The post is the only source for this table.
| Step | Amazon says | Not stated |
|---|---|---|
| Choosing the targets | A person: the actor provided a list of targeted IP addresses. | How the list was chosen. |
| Setting the goal | A person: variations of the goal "obtain shell access". The operator's prompts were in Simplified Chinese. | Whether the goal was changed during a run. |
| Each next step | The agents: planned, executed and adjusted "without the operator choosing each next step". | How many prompts or interventions each operation had. |
| The two failed operations | The agents: "nearly a full day without human direction" across the two. | Whether a person watched, restarted or stopped either. |
| Finding the flaw | The agents: identified the framework, mapped the API, found and tested the parameter. | Whether a person took part in that session. |
| A second university network | A person: "the threat actor stepped in", gave the agents a proxy into it, and directed them at other machines there. | How many machines, and what was done. |
| The share of the work done by agents | Nothing: no figure is printed. | The real autonomy fraction. |
So the post supports a narrower claim than its headline. A person chose the targets and the goal, wrote the prompts and, in one session, stepped in and gave the agents a route into a second network. The agents did the step by step work inside each operation, including the one that found the flaw, and Amazon says two of the three operations ran for "nearly a full day without human direction". Amazon also says that for the second university "we infer the actor had used the zero-day to compromise it": the compromise of that machine is Amazon's inference, not an observation.
A different documented case shows what the human layer can look like. Gambit Security's report of 22 September 2026 on a financially motivated operator against online retailers says the operator typed "1,951 prompts typed by the human across 260 sessions", "only a few prompts per target", into an orchestrating agent that launched Cairn and two other open source tools. That is a different operator on Gambit's account, and neither page says the two are connected.
The application Amazon does not name, and the records that do
The post says the target was "an open-source medical imaging AI application" and that the maintainers of the affected project were told. It prints no project or product name. Two records published in the same week name one. The project's advisory, GHSA-qc22-rmgc-vg94, is for MONAI Label, published 18 November 2026 and rated Critical. The CVE record published two days later lists that advisory and its fix pull request as references. The first figure sets the three pages side by side, with dashed cells where a page is silent.
Whether the three describe one flaw is an inference, so the match is laid out feature by feature, with what differs shown as plainly as what agrees.
Amazon's description set against the MONAI Label advisory and its pull request, as read on 21 November 2026. The last column is this site's reading, an inference.
| Amazon's post says | The MONAI Label pages say | This site's reading |
|---|---|---|
| An open-source medical imaging AI application. | The pull request's test list names radiology, pathology and endoscopy apps, and its text mentions DICOM and NIfTI. | Consistent with a medical imaging tool. |
| An upload endpoint with a filename parameter, tested for path traversal. | The upload endpoints take a caller-controlled image query parameter used as the image ID without validation. CWE-22. | Close. The parameter is called a filename by Amazon and an image ID by the advisory. |
| Arbitrary file writes on the server. | "An arbitrary file write can lead to remote code execution." | Matches. |
| A root reverse shell after a training API endpoint is triggered. | No training endpoint is mentioned. The advisory gives other examples of how a file write becomes code execution. | Not shown by the advisory. Amazon describes one route; the advisory lists different ones as examples. |
| No advisory or fix existed at discovery (10 September). | The fix pull request opened on 12 October; the advisory is dated 18 November. | Consistent in order. |
| The maintainers were notified by Amazon. | "Originally reported by @aws-sec-vdp" in a private advisory in the MONAI repository, which this site did not read. | Consistent. The handle reads as an Amazon security account; the page does not say who runs it. |
| The target's authentication is not described. | "Authentication is disabled by default"; with it enabled, the ANNOTATOR role is sufficient. | Amazon says nothing on whether the victim ran with authentication. |
| Amazon's post is dated 20 November. | A comment on the pull request of 16 October says the reporter needs it before 20 November. | Timing is consistent. Neither page says why. |
Two limits on that table. No saved page links the two sides in words: the advisory does not mention Amazon's campaign and Amazon's post does not mention the advisory. And one of the CVE record's three references is a press report on the Amazon findings, judged from its address as printed; this site did not read it. What would settle the question is Amazon naming the project. Until then the sentence this site can support is that the description fits, not that it is the same flaw.
Fixed, merged, released: three states of one fix
Amazon's post says the maintainers "they released a fix and a GitHub Security Advisory". The project's advisory says something narrower: patched versions "None", and "The fix will be included in the next MONAI Label release." The pull request the advisory points to says more. As saved at about 08:48 UTC on 21 November 2026, it was opened on 12 October 2026, approved by a reviewer on 17 October, and merged into the main branch on 19 October with 18 checks passed. So the fix has been in the project's main branch for 30 days at the date of the advisory (32 at the date of the CVE record), and in no release the advisory lists. Whether a release has been tagged since 18 November is not on any saved page.
The order matters. Amazon identified the Cairn server on 10 September. The fix pull request opened 32 days later and merged 39 days later (derived). The advisory followed 30 days after the merge. The pull request page prints a first commit message and a test file name that describe path traversal, and its title was shortened on the day it opened, dropping the words "to prevent path traversal"; the page does not say why. A comment on 16 October asks maintainers to coordinate with the reporting account, which it says has "time-sensitive work" and needs the change "before Nov 20th". Amazon's post is dated 20 November. That agreement of dates is on the page; the reason for it is not.
For a reader running MONAI Label the practical state, as saved, is this. Versions up to 0.8.5 are affected and none is listed as patched. The fix is public and merged but unreleased on the advisory's account. The pull request carries more than the security change: its own summary lists repairs to the project's continuous integration, new dependency declarations and test changes, including dropping Python 3.9 from the build matrices. Anyone taking the fix from the main branch rather than waiting for a release takes those with it (this site's observation from the page, not advice against it).
Google's threat tracker, covered in an earlier briefing, says threat actors are using agentic tools to reverse engineer patches by differential analysis. Here the order was the reverse: exploitation first, public fix second. Whether anyone used the public pull request before the advisory is not stated on any saved page.
The CVE record says exploited in the wild; CISA's rating on the same record says none
The CVE record for CVE-2026-133968 names mitre as the assigner. It was reserved at 18:31:37 UTC and published at 18:31:38 UTC on 20 November 2026, under a second apart (derived: 0.929 seconds), and last updated at 21:46 UTC. Its description reads that MONAI Label "through 0.8.5 allows directory traversal, with resultant remote code execution" via the image parameter of its datastore upload endpoint, "as exploited in the wild starting in September 2026". It scores CVSS 3.1 9.8 CRITICAL, with attack vector network, attack complexity low and privileges required none; that score was assigned by mitre, the record's assigner. It lists versions through 0.8.5 as affected, with a default status of unknown, and prints no fixed version. It prints the vendor as "Project-MONIA", a spelling that differs from the project's own name, and the weakness as CWE-24 where the advisory prints CWE-22 (both are traversal classes).
The same record carries a CISA-ADP entry stamped 21:34 UTC on 20 November, 3 hours 2 minutes 28 seconds after publication (derived), in the role "CISA Coordinator": Exploitation: none, Automatable: yes, Technical Impact: total. CISA-ADP is not the vendor and not the assigner; the rating is CISA's. The record's text says exploited in the wild from September 2026; CISA's rating says none. The two sit in one record and are a difference this site cannot resolve. The record names no source for the exploited claim, and the only account of this flaw being exploited on any saved page is Amazon's. This site did not read CISA's catalogue of known exploited vulnerabilities and does not state whether the entry is in it.
The GitHub Advisory Database entry for the same CVE, GHSA-2f43-rwqm-6h67, was published on 20 November 2026 and is marked Unreviewed, with affected and patched versions both Unknown. It repeats the record's description and the 9.8. The project's own advisory, as saved at about 08:48 UTC on 21 November, about 14 hours after the CVE was published (derived: 14 hours 16 minutes), still reads "No known CVE". Its text says the issue was "Solved here (#1915) and re-reported to generate CVE in the correct repository", and the CVE record names a request form as its generator; who asked for the CVE is not stated.
One more reading of the score, flagged as this site's. The 9.8 carries privileges required: none. The advisory says "Authentication is disabled by default" (the setting is named there), so in the default configuration any network client that can reach the server can exploit the flaw; with authentication enabled, "the ANNOTATOR role is sufficient". The 9.8 therefore reads as the score for the default configuration, not for a server with authentication on.
Three labels that do the work of evidence
A penetration testing framework. Amazon says Cairn "was built as a research and problem-solving engine for defenders". Gambit calls it "Cairn is an autonomous penetration testing engine.". Google's tracker calls it an open source, multi-agent pentesting framework. The label describes what it was built for. The saved pages describe it used against universities (Amazon), online retailers (Gambit) and, in Google's words, "GTIG observed CAIRN used in intrusion activity against targets including in the aerospace sector in the US". The Record quotes a Gambit researcher saying "it has no internal structuring for cybersecurity". A defender's name for a tool is not a control on who runs it, and none of the saved pages describes one.
Zero-day. Amazon's basis is an absence of any prior public record, at a stated confidence of high. That is a reasoned inference, and Amazon labels it as one by giving a confidence level. It is also a statement about public records on 10 September, not about who else knew.
Authentication off by default. The advisory says the default configuration is open to any client that can reach the server. Amazon found the attacker's own Cairn server "running without authentication" too. The same missing control appears at both ends of the story: it exposed the operator's server and, on the advisory's account, the default configuration of the victim's application.
What the sources state, and what they do not
Questions a reader would ask, against what the saved pages print as read on 21 November 2026.
| Question | Stated | Not stated |
|---|---|---|
| Which application was exploited? | Amazon: "an open-source medical imaging AI application". | Its name. The MONAI Label match is this site's inference. |
| Which versions are affected? | The advisory and the CVE record: through 0.8.5. | Whether Amazon's victim ran one of them. |
| Is there a patched release? | The advisory: patched versions None; fix in the next release. | A release date, or whether one exists since 18 November. |
| Was it exploited? | Amazon: yes, by the actor it describes. The CVE text: exploited in the wild from September 2026. | Anyone other than Amazon's actor. CISA-ADP: Exploitation none. |
| How many victims? | Universities; a second university network. | A count, the names, the country. The Record: AWS declined to name them. |
| Was data taken? | Nothing. | Whether any was, and what the actor did after the shell. |
| How autonomous was it? | Agents did the step by step work; a person set targets, goals and one proxy. | The fraction. |
| Which language model drove the agents? | Nothing on the Amazon page. | The model. Gambit names one for a different operator. |
| How many Cairn servers exist? | Nothing. | A count. |
Where the sources differ
Differences between the saved pages, unresolved.
| Point | One page says | Another says |
|---|---|---|
| The fix | Amazon: the maintainers "they released a fix and a GitHub Security Advisory". | The advisory: patched versions None. The pull request: merged to main on 19 October. |
| Exploitation | The CVE record text: exploited in the wild starting in September 2026. | The CISA-ADP entry on the same record: Exploitation none. |
| The second university | Amazon: "we infer the actor had used the zero-day to compromise it". | The Record: "The attacker then used the same apparent zero-day to breach a second university." |
| How many universities | Amazon: universities, one Redis university, a GPU server university, and a second university network; no total. | The Record: breaching "several universities in Asia". |
| The weakness class | The advisory: CWE-22. | The CVE record: CWE-24. |
| The parameter | Amazon: a filename parameter. | The advisory: an image query parameter used as the image ID. |
| Google's date | The saved Google page prints November 16, 2026, a Monday. | The Record says Google's report came "on Tuesday". |
The article follows each primary for what it states and keeps the disagreement where it is. The Record is a secondary: where it says more than Amazon, as on the second university, the article reports Amazon's wording and says the secondary goes further.
Context from Gambit, Google and The Record
Gambit wrote on 22 September 2026 about "a financially motivated operator" running three open source AI harnesses against online retailers, of which Cairn "for autonomous end-to-end exploitation" is one. It says 105 attack projects were launched between 10 and 15 September and that at least 27 companies were compromised to varying degrees. It says Cairn "then runs for hours until it achieves the objective, times out, or is stopped" and that "DeepSeek v4.1 Flash" was used in the Cairn attacks. Amazon names no model. Gambit's window of 10 to 15 September overlaps the day Amazon says it found the university campaign's server, 10 September. Neither page says the two actors are linked; Amazon calls Gambit's case "separate". Gambit sells a security platform, so its interest is of the same kind as Amazon's.
One observation, labelled as this site's. Amazon's attribution rests on Simplified Chinese prompts, China Standard Time timestamps and the target set. Gambit's page also prints Chinese language prompts, for an operator it calls financially motivated. The saved pages do not say language alone separates espionage from crime, and Amazon states its own attribution at moderate confidence.
Google's AI threat tracker, dated 16 November 2026, says it "independently observed" Cairn "deployed in the wild" and describes it as an observe, orient, decide, act loop in which specialised agents share a blackboard. On targets it names the aerospace sector in the US, and it names no actor, victim or date for that. The Record's article, a secondary dated 20 November, adds a quotation from the CISO of Amazon Integrated Security on the speed of the agents and reports that "AWS declined to name the universities or the country they are in".
The project's workarounds, as the advisory prints them
The advisory gives three workarounds for "until you can upgrade". They are the only mitigation text on any saved page for this flaw. The first line changes a setting that the advisory says is off by default. Amazon says its reverse shell ran as root; the third line is the one that bears on that (this site's reading). Copied exactly from the saved advisory.
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 with write access limited to the datastore directory.
What to do, in order
Take this with you
In the order worth doing (items 1 and 3 follow the project and Amazon; the rest are this site's suggestions)
- MONAI Label operators: check whether any server runs version 0.8.5 or earlier. The advisory lists no patched version, so apply its three workarounds now: turn authentication on, withhold the annotator role from untrusted users, keep the server off untrusted networks.
- Run the server as an unprivileged user whose write access is limited to the datastore directory, so a file write cannot reach application code, start up files or scheduled jobs.
- If you take the fix from the main branch before a release, read pull request 1915 first: it changes continuous integration, dependencies and tests as well as the datastore paths.
- Find research and medical imaging AI servers reachable from outside your network without authentication, including annotation and labelling tools in university and hospital research groups.
- Alert on high volume credential testing and long probing from one source, and wire the high confidence detections to automated response, as Amazon advises, with a person reviewing outcomes.
- Time your own triage: Amazon's agents ran 16 hours without a break. Ask how long an alert from the first probe waits before a person reads it.
- Do not wait for a CVE. Here the project advisory came two days before the CVE record and 71 days after the campaign began; a public fix pull request was merged 30 days before the advisory.
The question Amazon leaves open
Amazon told the maintainers, the maintainers merged a fix on 19 October, and the first public statement that the flaw existed came on 18 November, two days before Amazon wrote about the agents that used it. The pages show what happened in that order and not why. If an unauthenticated server and a public fix pull request were enough for a defender to act on 19 October, which of your own AI and imaging servers would have been on the list, and who would have told you?
Key facts
Sources
- PrimaryAmazon Threat Intelligence, Chinese-nexus actor's Cairn AI agents found a zero-day on their own, 20 November 2026 17:30 UTC, saved 21 November 2026 about 08:43 UTC; two addresses redacted by this site and not printed in the articleAmazon Web Services (AWS Security Blog)accessed 2026-11-21
- PrimaryPath traversal in datastore upload endpoints allows arbitrary file write, published 18 November 2026, saved 21 November 2026 about 08:48 UTC: affected versions, patched versions None, workarounds, the fixMONAI Label project (GitHub security advisory GHSA-qc22-rmgc-vg94)accessed 2026-11-21
- PrimaryThe fix pull request, saved 21 November 2026 about 08:48 UTC: opened 12 October, merged to main 19 October 2026, the change summary and the comment of 16 OctoberMONAI Label project (GitHub pull request 1915)accessed 2026-11-21
- PrimaryCVE record JSON, assigner mitre, published 20 November 2026 18:31 UTC, updated 21:46 UTC, with the CISA-ADP SSVC entry; saved 21 November 2026 about 08:48 UTCCVE Program (CVE record CVE-2026-133968)accessed 2026-11-21
- PrimaryAutonomous AI agents and online retailers, 22 September 2026: a financially motivated operator, Cairn, Strix and Hermes; read for its own claims about Cairn, saved 21 November 2026 about 08:48 UTCGambit Securityaccessed 2026-11-21
- PrimaryGTIG AI Threat Tracker, 16 November 2026: the Cairn paragraph; copy saved 20 November 2026 about 15:19 UTCGoogle Threat Intelligence Groupaccessed 2026-11-21
- Reported byUnreviewed entry for CVE-2026-133968, published 20 November 2026, saved 21 November 2026 about 08:48 UTCGitHub (GitHub Advisory Database GHSA-2f43-rwqm-6h67)accessed 2026-11-21
- Reported byAmazon says Chinese hackers used Cairn AI tool to discover apparent zero-day during attack on university, 20 November 2026; a secondary, saved 21 November 2026 about 08:42 UTCThe Record from Recorded Future Newsaccessed 2026-11-21


