Amazon's threat intelligence team has captured records of AI agents running a campaign to steal API keys for AI services, then using the stolen keys to pay for the very models powering their hunt for more keys

1

3

2

. The loot pays the compute bill: the records describe an attack funding itself.

Source: VentureBeat

Source: VentureBeat

The records came from Amazon's MadPot decoy-sensor network, which posed as popular AI model APIs and collected what the agents sent it between October 3 and November 5

1

3

2

. At moderate confidence, Amazon judges that Artex, an open-source AI penetration-testing framework, is what the agents ran on

1

3

2

. It also names separate China-linked groups as the operators, pointing to Simplified Chinese prompts and China time-zone settings

1

3

2

. The records capture what the Artex agents were pursuing; Amazon never saw the attacks directly and cannot check the results the agents claimed

1

3

2

.

How the campaigns worked

In the main campaign, the records describe agents scanning for AI applications left open to the internet and pulling out the provider keys stored inside

1

3

2

. The agents checked that each key worked, then added the working ones to a shared pool

1

3

2

. Because one weak admin page can expose credentials for many providers at once, the agents prioritized apps that hold keys to several services

1

3

2

. Named targets in the Artex campaign records include gateway tools such as One API, New API and LiteLLM, and AI app builders including Dify, FastGPT, Langflow and Open WebUI

1

3

2

. Being named as a target does not mean each was successfully compromised

1

.

Source: Hacker News

Source: Hacker News

The Artex playbook matched each named product to known entry points, from default admin credentials to two named security flaws

1

3

2

. The agents also hunted for keys leaked in public places, sweeping GitHub repositories, npm and PyPI packages, container images and paste sites

1

3

. Where they found cloud credentials, the records describe the agents prioritizing high-end AI models over general cloud access

1

.

The records behind the Artex assessment also hold a second campaign, in which agents used deepfakes to try to get past know-your-customer checks at two crypto exchanges

1

3

. One verification vendor's server rejected every recorded video attempt, and Amazon did not say whether the fabricated identity-document photos passed document checks

1

3

.

A tool that won't go away

Artex's developer took the GitHub repository down on October 8 amid abuse reports

1

3

2

. That did not matter much: copies on community forks, in archives and in a paid version sold as Artex Pro keep it in circulation

1

3

2

. By early November, the Federal Bureau of Investigation counted more than 1,600 Artex consoles left open to the internet

1

3

2

. Cisco Talos dates the framework's publication to July and says it is used largely by financially motivated criminals and China-nexus actors

3

.

The bureau issued a flash alert on November 13 saying attackers of many kinds have used Artex since at least August, across sectors worldwide

1

3

2

. Some intrusions, it says, reached privileged access in under three hours

1

3

2

. The alert draws on 11 earlier reports and covers no fresh incident of its own

1

. Its examples include an earlier hit on South Korean financial firms that took personal data on about 67,000 people

2

.

The wider wave

Other vendors have been documenting the same tooling. Cloudflare incident responders handled multiple Artex-related investigations over the past eight weeks

1

. In one case, a compromised account-level token with no restrictions was used to fire 400 database dump requests in an hour

1

. Cloudflare says the common root cause in more than one incident was static tokens that never expired and granted broad access across the customer's cloud setup

1

. Cloudflare has not disclosed whether the dump requests returned data, how the token was first stolen, or whether the attackers kept their foothold

1

.

Microsoft has separately reported Artex use for GitHub secret hunting, plus sign-in attempts against more than 2,250 downstream Microsoft 365 tenants

1

2

.

CrowdStrike's Adam Meyers said a human operator might send a few dozen commands an hour at a target

2

. The most CrowdStrike had seen from an agent, Meyers said, was "like 1,500 commands in an hour"

2

. In the three days before its post, Cloudflare counted more than 50 families of AI penetration-testing frameworks trying to scan its customers

1

3

. Cisco Talos links more than 2,500 addresses to Artex in its indicator files

1

3

.

Amazon warned Amazon Web Services customers it judged potentially affected and published indicators of compromise from the Artex research, including 11 IP addresses from the decoy network

1

3

2

. The bureau cautions against blocking those addresses on the advisory alone, since cloud addresses may have been reassigned

2

. Its advice for defenders is to treat AI apps, model gateways and inference endpoints as production systems, assume exposed secrets will be found quickly, and watch for unexpected spending

1

3

. "None of the techniques in these campaigns were new," Amazon's threat intelligence team wrote. "What AI agents change is the speed and persistence of the attacker."

1

3

Source: Quartz

Source: Quartz

© 2026 TheOutpost.AI All rights reserved