Amazon: AI Agents Autonomously Discovered a Zero-Day Vulnerability in 11 Minutes
AI agents run by a China-linked threat actor identified a previously unknown vulnerability in an open-source medical imaging application and exploited it on their own, Amazon’s security blog reported on 20 November 2026.

· Originally published by ontime+ · Last verified: 22 Nov 2026

Key Points
- Amazon says AI agents discovered an unknown flaw and exploited it with minimal human direction.
- The operator gave the agents one goal — shell access — and a list of IP addresses.
- Agents moved from scan to arbitrary file write in roughly 11 minutes.
The latest:
AI agents run by a China-linked threat actor identified a previously unknown vulnerability in an open-source medical imaging application and exploited it on their own, Amazon’s security blog reported on 20 November 2026. The write-up, by Amazon Chief Security Officer CJ Moses, says the human operator supplied only a target phrase and a list of IP addresses before the agents took over.
Details:
- The discovery: Amazon’s threat intelligence team found a publicly exposed server on 10 September 2026 running an instance of the Cairn framework with no authentication, according to the report. The full log was visible: project configurations, operator commands and AI output. Commands were written in simplified Chinese and script timings were set to China time.
- The framework: Cairn is an open-source autonomous penetration-testing framework built for researchers and defenders, the report says. It accepts a goal in natural language and coordinates multiple agents to reach it. Here the stated objective was shell access against a supplied target list.
- The campaign: The operation ran roughly 48 hours starting 8 September 2026 against IP addresses belonging to East Asian universities outside mainland China, according to Amazon. Indicators of compromise logged between 8 and 10 September include 142.4.98.145 and 142.4.107.34.
- What failed first: Over about 16 hours and 148 sequential actions, the agents tried known techniques against an exposed Redis server — replication abuse to load code, script sandbox escape, writing to configuration files, memory corruption — and attempted a Log4Shell injection on another service on the same system. All of it failed, Amazon said.
- The brute-force effort: The agents mapped the internal Docker network through an SSRF flaw in a neighboring Qdrant instance, then sprayed credentials at PostgreSQL, SSH and other services without success. On one university GPU server they worked more than 6 hours, building roughly 80 real usernames from faculty directories, lab sites, archived pages and academic papers.
- The scale: Those usernames fed more than 30,000 credential combinations tested against SSH, an admin panel, a storage device, a VPN server and the campus gateway, according to the report. Not one login succeeded. The zero-day path opened only after every conventional route was exhausted.
- The zero-day path: Against the AI-powered medical imaging application, the agents fingerprinted the framework, mapped API endpoints and found a file-upload point with a filename parameter. They tested it for path traversal and it was exploitable. From interface scan to arbitrary file write took roughly 11 minutes, Amazon said.
- Root access: In a later session the agents wrote a Python reverse payload into sitecustomize.py, then called a training endpoint that spawned a Python process importing the payload and returning a reverse shell with root privileges. The operator then supplied a proxy into a second university network running the same software.
- The attribution: Amazon assesses with medium confidence that the actor is China-linked, citing the simplified Chinese commands, China-time script scheduling and the profile of the targeted universities. It assesses with high confidence that the flaw was a zero-day, since no patch, report or prior database entry existed.
- The disclosure: Amazon notified the project maintainers, who issued a fix and a GitHub security advisory, GHSA-qc22-rmgc-vg94. Affected institutions, which are not AWS customers, were informed through a government partner. Gambit Security was among the first to document a group using Cairn, in a card-theft operation. Amazon advises defenders to prioritize high-confidence detections triggering automated response, with human oversight applied to outcomes rather than individual actions.
Between the lines:
The timeline is the finding. The agents burned roughly 16 hours and 148 actions on known techniques that all failed, then needed about 11 minutes once they located an unpatched upload parameter. Amazon’s recommendation to review outcomes instead of individual actions follows directly from that gap: a defender reviewing each of 148 steps would still be reading logs when root access arrived.
What’s next
Watch patch uptake on the advisory GHSA-qc22-rmgc-vg94, whether the affected East Asian universities report follow-on intrusions, and whether additional vendors document Cairn-based operations after Gambit Security’s card-theft case.