en
Back to the list

Amazon Says AI Agents Found and Exploited a Zero-Day

source-logo  thedefiant.io 1 h
image

A suspected China-linked attacker used AI agents to discover and exploit a previously unknown software flaw and take control of a university server, Amazon security chief CJ Moses said in a Nov. 20 AWS Security Blog post.

The affected application was MONAI Label, an open-source medical-imaging AI tool. The agents took about 11 minutes to progress from mapping the application's interfaces to writing files on the server, according to Moses, chief information security officer of Amazon Integrated Security. In a later session, they used the flaw to establish remote command access running as root, giving them administrative control.

The account describes AI handling vulnerability discovery as well as attack execution—not merely reusing an exploit for a known bug. Amazon assessed with high confidence that the flaw was a zero-day, citing the absence of a corresponding fix, issue, advisory or prior vulnerability-database mention at the time of discovery.

Know when your
coins move

Alerts, real-time prices, and market news — all in one app
qr-code Point your phone camera to install
4.8 based on 40K reviews in the App Store and Google Play

Amazon said it found the evidence on Sept. 10 in an attacker-operated instance of Cairn, an open-source AI penetration-testing framework, left publicly accessible without authentication. Its exposed history included operator prompts and AI output. The campaign targeted universities in East Asia outside mainland China over approximately 48 hours beginning Sept. 8.

Amazon assessed the China link with moderate confidence, citing Simplified Chinese prompts, script timestamps configured to China Standard Time and targets it considered likely to be of strategic interest to China.

What the Agents Did Without an Operator

The human supplied target IP addresses and variations of the goal “obtain shell access,” Moses wrote. The agents then planned, executed and adjusted their approach without the operator selecting each next step.

Two unsuccessful operations, against Redis and a university GPU server, underpinned Moses's statement that “the agents kept adapting for nearly a full day without human direction.” That description referred to persistent failed attempts, not a day spent exploiting the zero-day.

Persistence Was Not the Same as Success

The campaign was not entirely hands-off. After the zero-day's discovery, the operator supplied a proxy into a second university network and directed the agents to target other machines through it. Amazon inferred, rather than directly established, that the attacker had used the same flaw to compromise the server providing that tunnel.

The maintainer advisory, published Nov. 18, describes a path-traversal flaw: uploaded files could be written outside their intended storage directory, potentially allowing remote code execution.

The advisory lists versions through 0.8.5 as affected. It says the code fix is available but will be included in the next release, and lists no patched version. Until an upgrade is possible, maintainers recommend enabling authentication, keeping the server off untrusted networks and running it without administrative privileges.

thedefiant.io