---
title: "Amazon decoy servers catch AI agents stealing API keys; FBI clocks some intrusions at under three hours"
description: "Amazon's decoy servers recorded AI agents hunting provider keys to fund their own inference. KERNEL's November 4 notice shows a stolen key opening four of its GPU servers, and six days later the breached vendor started selling the one-use credential that would have stopped it."
publisher: "The Inference"
section: "Security"
published: 2026-11-16T15:19:27.741Z
modified: 2026-11-16T15:19:27.741Z
canonical: https://theinference.org/article/amazon-decoy-servers-catch-ai-agents-stealing-api-keys-fbi-clocks-some-intrusions-at-under-three-hours
language: en
keywords: "Security, AI Agents, API Keys, Amazon, KERNEL"
---

# Amazon decoy servers catch AI agents stealing API keys; FBI clocks some intrusions at under three hours

> Amazon's decoy servers recorded AI agents hunting provider keys to fund their own inference. KERNEL's November 4 notice shows a stolen key opening four of its GPU servers, and six days later the breached vendor started selling the one-use credential that would have stopped it.

[The FBI's flash alert on ARTEX](https://casrai.org/news/fbi-artex-ai-attack-framework-warning-2026), published November 13, says criminals have used the open-source, LLM-driven framework since at least August to automate intrusions, in some cases from first contact to privileged access in under three hours [1]. Three days later, on November 16, [Amazon](https://theinference.org/markets/companies/amazon)'s threat intelligence team published records from decoy servers showing AI agents sweeping exposed AI apps for provider keys; Amazon has only moderate confidence the agents ran on ARTEX [2].

The two bulletins share one lesson for anyone selling agent infrastructure: [the API key that powers your customers' agents](https://theinference.org/article/sap-patches-a-10-out-of-10-flaw-that-can-send-ai-agents-credentials-to-attackers-earnings-due-october-21) is now the attackers' first target, and who holds it safely matters more than what the agent can do.

## Attackers hunt keys, not just networks

The mechanism is plain. Autonomous agents need inference, inference needs keys, and keys sit inside AI apps built for humans. [The records Amazon captured](https://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.html) show agents exploiting exposed LiteLLM, Dify and One API installs, extracting provider keys into a shared pool, and using the stolen keys to run the models that searched for more keys [2].

The FBI found Chinese-origin models such as [DeepSeek](https://theinference.org/markets/companies/deepseek) most prevalent on attacker-controlled ARTEX instances [1]. [Cloudflare](https://theinference.org/markets/companies/cloudflare)'s responders watched one static CI/CD token, with no expiry or IP limits, draw 400 database dump requests in an hour [2]. What agents change, as Amazon put it, is the speed and persistence of the attacker; the flaws were ordinary [2].

## KERNEL is the breach that proves the bulletin

[KERNEL](https://theinference.org/markets/companies/kernel) is the one vendor that has tied the tool to itself. The San Francisco company supplies browser infrastructure for AI agents and says more than 4,000 companies use it [5][6]. [Its November 4 notice says](https://tribeai.com.br/en/articles/kernel-security-incident-compromised-api-key-logging-provider-gpu-servers/) a stolen KERNEL API key probed access on October 24, reached a staff login at its logging provider on October 27, and on November 3 opened four GPU servers, three running customer browser sessions; a cryptominer started on a development server [3][4].

The company says the breach touched 0.28% of its customers, and [its evidence points to ARTEX as the campaign's command center](https://ai.info/news/kernel-secret-exposing-breach-ai-assisted-campaign), though that attribution is KERNEL's own belief [3][4].

| Date | What KERNEL's notice says happened |
| --- | --- |
| October 24 | A stolen KERNEL API key probes access |
| October 27 | Reaches a staff login at KERNEL's logging provider |
| November 3 | Four GPU servers opened, three running customer browser sessions; a cryptominer starts on a development server |
| November 4 | KERNEL publishes its incident notice |

*The sequence in KERNEL's November 4 incident notice, as reported by TribeAI and AI.info on November 5 and 6.*

## The breached vendor now sells the vault

Six days later KERNEL shipped agent payments with [Stripe](https://theinference.org/markets/companies/stripe) and Agentcard: the credential is issued one-use, bound to merchant and amount, and never enters the developer's application or the agent's model context [5][6]. That is the custody pattern the FBI's own advice points toward, now sold as a product, and it answers the exact failure KERNEL itself suffered: a standing key that opened doors across systems [3][5].

## Stripe is the listed way to own the custody trade

The listed expression here is Stripe, whose Link issues one-use credentials into KERNEL's vault and already sits inside the custody pattern agents need [5]. Vendors whose product is a standing API key and an untrusted agent could face higher customer costs for rotating, scoping and vaulting keys, which may weigh on how that private market prices them; no source prices these companies, and the connection between the FBI alert and KERNEL's notice is the inference of this article, not a claim any party has made.

What to watch next: whether the next major agent-infrastructure breach, if one comes, is traced to a vaulted one-use credential. If it is not, custody wins the argument.

## Takeaway

Key custody, not model power, is now the agent economy's trust layer: a standing API key carried attackers to KERNEL's GPU servers, and one-use credentials are the product KERNEL now sells.

## Sources

1. [CASRAI, FBI Warns of ARTEX, an AI Framework That Automates Intrusions, 14 November 2026](https://casrai.org/news/fbi-artex-ai-attack-framework-warning-2026)
2. [The Hacker News, Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys, 16 November 2026](https://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.html)
3. [TribeAI, Kernel reports security incident involving a compromised API key, logging provider access and 4 GPU-enabled servers, 5 November 2026](https://tribeai.com.br/en/articles/kernel-security-incident-compromised-api-key-logging-provider-gpu-servers/)
4. [AI.info, Kernel Believes Attackers Are Leveraging ARTEX in a Larger Cryptomining Campaign, 6 November 2026](https://ai.info/news/kernel-secret-exposing-breach-ai-assisted-campaign)
5. [Business Wire, KERNEL Lets Browser Agents Pay on Website Checkouts Without Exposing Card Data, 10 November 2026](https://www.businesswire.com/news/home/20261110606274/en/KERNEL-Lets-Browser-Agents-Pay-on-Website-Checkouts-Without-Exposing-Card-Data)
6. [Digital Transactions, The AI Upload: Kernel's AI Tool Limits Card Data Exposure; 40% of Young Consumers Will Employ AI in Gift Shopping, 12 November 2026](https://www.digitaltransactions.net/the-ai-upload-kernels-ai-tool-limits-card-data-exposure-40-of-young-consumers-will-employ-ai-in-gift-shopping/)


---

Amazon decoy servers catch AI agents stealing API keys; FBI clocks some intrusions at under three hours — The Inference. Canonical: https://theinference.org/article/amazon-decoy-servers-catch-ai-agents-stealing-api-keys-fbi-clocks-some-intrusions-at-under-three-hours
