Three AI Servers Landed on Exploited Lists Today: MONAI Label Has No Fixed Release, 9Router Lets Strangers Spend Your LLM Credits, and Ray's 2023 File-Read Bug Is Still Being Hit

If you run MONAI Label, 9Router or Ray on a network anyone else can reach, take it off that network today. All three were flagged as exploited on Friday, November 20, 2026. None of them is in CISA's KEV catalog yet, so a KEV-driven patch queue will not show them to you.
What to do, by product
- Who is exposed: every monailabel version through 0.8.5, which is still the newest release on PyPI. The GitHub advisory lists the vulnerable range as "<= 0.8.5" and leaves the patched version empty. Its exact words: "The fix will be included in the next MONAI Label release."
- The fix is pull request 1915, "fix: confine datastore write paths", merged into main on October 19, 2026 (merge commit 113527bf). If you can run unreleased code, build from main at that commit or later. Otherwise, use the advisory's three workarounds:
- Turn on authentication with MONAI_LABEL_AUTH_ENABLE=True. It is off by default. The advisory says "in the default configuration any network client that can reach the server can exploit this," and with authentication on, the ANNOTATOR role is still enough, so give that role to no one you do not trust.
- Do not expose the server to untrusted networks. By default the server listens on all interfaces (host 0.0.0.0) on port 8000.
- Run it as an unprivileged user that can write only to the datastore directory. In the attack Amazon described, the shell ran as root.
- The flaw: if requests reach Next.js without the bundled custom-server.js wrapper, the app trusts a client-supplied X-9r-Real-Ip header. Sending "X-9r-Real-Ip: 127.0.0.1" makes the attacker look local, and the API-key check on /api/v1/* is skipped. The attacker can then spend your paid LLM credits and list the providers and models you have set up.
- The sources disagree on the fixed version. NVD and VulnCheck say "fixed in version 0.5.6". The GitHub advisory says vulnerable "<= 0.5.4" and first patched "0.5.8". npm has no 0.5.6: the registry jumps from 0.5.4 (June 18) to 0.5.8 (June 21). So 0.5.8 is the first patched version you can actually install from npm.
- If you ran an old version exposed to the network, rotate every provider API key you configured and check each provider's billing for usage you did not make. The point of this bug is spending someone else's inference.
- The flaw: an unauthenticated local file read through the dashboard's /static/ directory. NVD: "LFI in Ray's /static/ directory allows attackers to read any file on the server without authentication." The GitHub advisory gives the vulnerable range as "< 2.8.1" and the fix as 2.8.1.
- Upgrading does not make an exposed Ray cluster safe. In its November 2023 post, Anyscale says "you must prevent access to your Ray cluster from untrusted machines (e.g., the public internet)." The dashboard and Jobs API have no authentication by design, which is the separate, disputed CVE-2023-48022 behind the ShadowRay campaigns. If port 8265 is reachable from the internet, version 2.8.1 still lets people submit jobs.
Am I affected? A self-check
- Version: run "pip show monailabel". A version of 0.8.5 or lower, installed from PyPI, is vulnerable.
- Process and port: look for a monailabel start_server process or a uvicorn process serving monailabel.app:app, listening on port 8000. Check whether that port can be reached from outside your network.
- Signs of compromise: look in the access logs for PUT requests to /datastore/image or /datastore/label whose image or tag value contains "..", "%2e" or a slash. Look for a sitecustomize.py in your Python site-packages that you did not put there, or one changed recently. Amazon says the attackers wrote their reverse shell into sitecustomize.py and then called a training endpoint to start a new Python process that loaded it. Also look for files written outside the datastore directory, and for connections to the two addresses Amazon published, listed below.
- Version: run "npm ls -g 9router", or check package.json if you deployed from source. Any version below 0.5.8 is vulnerable.
- Port: 20128 by default, for both the dashboard and the OpenAI-compatible API at /v1.
- How it was started: the advisory says the danger is running Next.js directly, without custom-server.js. The README's from-source steps start the app with "npm run dev" or "npm run start" using HOSTNAME=0.0.0.0. Check which one you used.
- Signs of misuse: look for requests from non-loopback addresses that carry an X-9r-Real-Ip header, a 200 on /api/v1/models from an address you do not recognize, and provider usage you cannot explain.
- Version: run "ray --version". Anything below 2.8.1 has the file-read bug.
- Port: the dashboard listens on 8265 by default. Check whether it can be reached from outside your network.
- Signs of probing: look for dashboard requests under /static/ that contain "..", "%2e" or "..;", and for job submissions you did not make.
What is confirmed
- MONAI Label was exploited as a zero-day by AI agents. This comes from Amazon Threat Intelligence, in an AWS Security Blog post by CJ Moses dated November 20, 2026. Amazon found a public, unauthenticated instance of Cairn, an open-source autonomous AI penetration-testing framework, on September 10. Its history showed a campaign against university IP addresses in East Asia outside mainland China, running for about 48 hours from Tuesday, September 8. Against the medical imaging server, Amazon says "In about 11 minutes, the agents went from mapping the application's API to arbitrarily writing files on the server," and later got a reverse shell "running as root." Amazon also infers that the operator used the same zero-day to compromise a second university server and turn it into a proxy. Amazon assesses "with moderate confidence" that the actor is Chinese-nexus, based on Simplified Chinese prompts, China Standard Time timestamps in the scripts, and the choice of targets.
- NVD's description says the bug was "exploited in the wild starting in September 2026." The CVE was published November 20 at 19:18 UTC and is rated CVSS 9.8, from the CNA (MITRE).
- VulnCheck KEV lists all three with date_added 2026-11-20. This is VulnCheck's data. VulnCheck's date_added is the date of its earliest exploitation evidence, not the date it added the entry. For MONAI Label, the evidence VulnCheck cites is the AWS post and the CVE record; it shows no canary sighting. For 9Router and Ray, VulnCheck's own canaries (its honeypots) recorded exploitation on November 20.
- None of the three is in CISA KEV. We checked catalog version 2026.11.20, released November 20 at 17:48 UTC, with 1,790 entries. It also does not contain CVE-2023-48022, the Ray no-authentication CVE. Note that the CISA SSVC entry NVD attaches to the MONAI CVE, timestamped 21:34 UTC on November 20, reads "exploitation: none." It conflicts with Amazon's report and with NVD's own description. For 9Router, SSVC reads "poc."
What is not known
- What Amazon's two addresses did. Amazon says only that they "appear in records dated September 8 to 10, 2026." It does not say whether they are the Cairn server, a staging host or exploit sources.
- What VulnCheck's canaries saw for 9Router and Ray: the volume, the source addresses, and whether the Ray hits were actual file reads or broad probing. We read the entry's flags and dates. We did not read the canary records themselves, and we did not republish them.
- Whether anyone besides Amazon's actor is exploiting MONAI Label. VulnCheck shows no canary sighting for it.
- When MONAI Label will ship a release containing the fix. Neither the advisory nor the PR gives a date.
- The Gambit Security report that Amazon cites about a separate, financially motivated Cairn operator. We did not read it.
How we know: our edge
We do not run any of these three products. Our edge sits in front of our own sites, so what follows is what scanners sent us, not a measurement of exploitation. The window is August 1 to November 20, 2026, and covers 700,997 edge blocks, 179,121 edge canary records and 5,776 origin honeypot captures. None of the matched paths were our own /post/ pages.
- MONAI Label: zero requests to /datastore/image or /datastore/label across all three instruments. A broad /info pattern matched 29 blocks and 3 canaries, but those were LiteLLM, Spring actuator and Symfony profiler paths, not MONAI. We do not count them. One limit: our edge records store the path but not the query string. The image parameter is in the query string, so we could see the endpoint being requested but not the traversal payload.
- 9Router: 7 blocks on /api/v1/models, from 6 addresses, between September 30 and November 19. That path is the generic model-list route of many OpenAI-compatible servers, so these are AI-API recon, not 9Router sightings. The thing that decides this CVE is the X-9r-Real-Ip header, and our edge records do not store request headers. For this CVE, our edge's answer is UNKNOWN, not zero.
- Ray: 167 blocks and 19 canaries matched traversal under /static/. Almost all of them were "/static/..;/.env" and "/static/..;/..;/..;/..;/proc/self/environ", requests for secrets files that also target Java and Spring path handling, and they arrived on our port 443, not Ray's dashboard port 8265. These are shared recon paths, not a Ray sighting, and we do not count them as one. Our edge cannot see port 8265, so for exploitation of a real Ray dashboard our answer is UNKNOWN.
- Public exploit code: our exploit harvester holds no PoC repository for any of the three CVEs. That is a statement about our harvester, not proof that no PoC exists. Huntr's report for the Ray bug is tagged "Exploit" on NVD.
Feed
Amazon published two addresses. We added both to the feed at confidence 80, source manual-batch-cairn-monai-label, attributed to Amazon and not presented as our own observation. Both read back by primary key, and our correlate lookup now returns held:true for 142.4.98.145. We confirmed the records are in our index. We have not yet seen the rows in a built ips.csv. Neither address had been in our corpus before today, and none of their /24 neighbors were.
Block these: get the free feed (docs: docs.dugganusa.com)
- 142[.]4.98.145 (Amazon, records dated September 8 to 10, 2026)
- 142[.]4.107.34 (Amazon, records dated September 8 to 10, 2026)
For 9Router and Ray: none published in the sources we read. The GitHub advisories, NVD and the VulnCheck entries name no indicators.
Block these: get the free feed
Sources
- Amazon Threat Intelligence (CJ Moses), "Chinese-nexus actor's Cairn AI agents found a zero-day on their own," AWS Security Blog, November 20, 2026: https://aws.amazon.com/blogs/security/chinese-nexus-actors-cairn-ai-agents-found-a-zero-day-on-their-own/
- MONAI Label advisory GHSA-qc22-rmgc-vg94: https://github.com/Project-MONAI/MONAILabel/security/advisories/GHSA-qc22-rmgc-vg94 and the fix PR: https://github.com/Project-MONAI/MONAILabel/pull/1915
- NVD, read through the NVD 2.0 API (the HTML pages render with JavaScript and came back empty to our fetcher): CVE-2026-133968, CVE-2026-56681, CVE-2023-6020
- 9Router advisory GHSA-5mj8-gf6m-fhw8: https://github.com/decolua/9router/security/advisories/GHSA-5mj8-gf6m-fhw8 and the npm registry version history for 9router
- Ray advisory GHSA-6cxr-8q3m-jwrr, and Anyscale, "Update on Ray CVEs," November 30, 2023: https://www.anyscale.com/blog/update-on-ray-cves-cve-2023-6019-cve-2023-6020-cve-2023-6021-cve-2023-48022-cve-2023-48023
- VulnCheck KEV, read through VulnCheck's API (VulnCheck's data, with attribution)
- CISA KEV JSON, catalog 2026.11.20: https://www.cisa.gov/sites/default/files/feeds/known_exploited_vulnerabilities.json
- The Register, November 20, 2026, which reports from Amazon's findings and is not an independent source
Our confidence in this write-up is high but capped at 95%. The parts we could not see are listed above.
Read next: CISA Added Synology DSM, Apache Axis and LiteLLM to Its Exploited List, another AI gateway that landed on an exploited list, with a version check for it.
Every indicator in this post is in the feed. Free.
1.58M+ IOCs, STIX 2.1 / TAXII, 88% novel vs ThreatFox, exploited-CVE leads ahead of CISA. No credit card — a free API key in 30 seconds, and you can audit every claim above against the live endpoints.