Amazon Says Chinese-Nexus Cairn Agents Found a Medical Imaging Zero-Day and Took Root
Amazon found Cairn used against East Asian universities. Agents discovered a MONAI Label path traversal, wrote a root shell, then used a second-campus tunnel.
Amazon's CISO published a campaign where the operator typed "obtain shell access" and left. The agents found a zero-day in a medical imaging app, wrote a root shell, and later got a human-built tunnel into a second university.
CJ Moses wrote on November 20 that Amazon Threat Intelligence found the work on September 10, 2026, on a publicly hosted instance of Cairn with no login. Cairn is an open-source autonomous pentest framework. Prompts were in Simplified Chinese. Script timestamps used China Standard Time. The campaign ran about 48 hours from September 8 against university IP addresses in East Asia outside mainland China.
The operator supplied a list of targets and a goal. The agents planned, ran, and revised the rest.
Sixteen hours of misses, then 11 minutes on a new bug
The longest job hit an unauthenticated Redis server at one university. Over about 16 hours and 148 sequential actions, the agents tried replication tricks, sandbox escapes, config writes, memory corruption, and a Log4Shell shot at a neighboring service. They mapped an internal Docker network through an SSRF in a co-located Qdrant instance and sprayed PostgreSQL, SSH, and other logins. Nothing landed.
A second job spent more than six hours on a university GPU cluster. The agents scraped faculty directories, lab sites, archives, and papers for about 80 real usernames, then tried more than 30,000 password combinations against SSH, an admin dashboard, storage, a VPN, and the campus portal. No login worked. Moses's point is that they kept adapting for nearly a day without a person picking the next step.
The third target was an open-source medical imaging AI application. The agents mapped its API, found an upload endpoint with a filename parameter, and tested path traversal. In about 11 minutes they went from mapping the API to writing files anywhere the process could write. In a later session they dropped a Python reverse shell into sitecustomize.py, which the interpreter imports at startup, then hit a training API so a new process would spawn. That process came up as root.
Amazon assesses with high confidence that this was a zero-day: no advisory, no patch, and no prior mention in vulnerability databases for that project at the time.
The operator came back with a tunnel
After the agents found the hole, the human handed them a proxy into a second university running the same software. Amazon infers the actor had already used the zero-day there. The agents were told to hit other machines through that tunnel.
Amazon assesses a Chinese-nexus actor with moderate confidence, on the language, the timezone, and the choice of universities. AWS notified the project's maintainers, who shipped a fix and a GitHub advisory. The victim schools were not AWS customers. AWS used a government partner to warn them.
MONAI Label's advisory, reported by AWS and published November 18, matches the hole: PUT /datastore/image takes a caller-controlled image parameter, writes through shutil.copy with no path check, and ships with authentication off. Versions through 0.8.5 are affected. The fix is pull request 1915, for the next release. Until then the project says to enable auth, keep the server off untrusted networks, and run it unprivileged.
Cyberpresso already covered a different Cairn operator who pointed agents at Deere, SharkNinja, and Cummins for about $50 a break-in. Moses points at the same class of tool in Gambit's retailer-theft reporting. Amazon's new case is the one where the agents found the bug themselves.
The addresses Amazon listed for September 8 to 10 are 142.4.98.145 and 142.4.107.34.
Some offers on this page may be paid placements or contain affiliate links.
Cyberpresso: daily cyber & AI brief
Free daily newsletter, read in 5 minutes.
Subscribe free