EN AR

Amazon decoys capture AI agents hunting API keys in LiteLLM, Dify

Tamara khadrah

Key Points

  1. Amazon says attacker-run AI agents scanned exposed AI apps and extracted stored provider keys from them.
  2. Stolen keys powered the very models that directed the search for more keys, the records show.
  3. Amazon cannot verify the agents' reported outcomes, and assesses China-linked operators with moderate confidence.

The latest:

Attacker-operated AI agents spent weeks probing internet-exposed AI applications such as LiteLLM, One API and Dify to pull out the model-provider keys stored inside them, Amazon’s threat intelligence team said this week, citing records its decoy servers captured. The team said it has moderate confidence the agents ran on ARTEX, an open-source AI penetration testing tool, and that stolen keys funded the models driving the hunt.

Details:

  • The trap: Amazon’s decoy network, MadPot, posed as popular large language model services. From October 3 until late last week, the agents sent it requests that often carried a campaign’s full history, including the operators’ instructions and the findings the agents reported back to their handlers.
  • The verification limit: The captured records describe work done on systems MadPot does not monitor. “Amazon Threat Intelligence did not observe that activity as it happened and cannot verify the outcomes the agents reported, but the records show what the actors were pursuing,” the team said.
  • The attribution: Amazon assessed with moderate confidence that the operators were separate China-linked groups, citing Simplified Chinese prompts and China time zone settings. The records cover hundreds of targets; Amazon notified AWS customers who may have been affected.
  • The playbook: The records list known entry points per product: default administrator credentials such as LiteLLM’s default master key, unfinished installs an outsider could take over, management endpoints needing no login, an authentication bypass via the isCliRequest header in One API family systems, and SSRF reaching cloud metadata services.
  • The target list: Named targets include gateways One API, New API, Veloera, Sub2API and LiteLLM, plus app builders Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. The agents prioritized apps holding keys for many providers, where one weak admin page exposes many accounts.
  • The hunting ground: To find leaked keys the agents used internet scanning services and certificate transparency logs, and searched GitHub repositories, event archives, npm and PyPI packages, container image layers, notebooks, model-hosting spaces, paste sites and Telegram channels. Found cloud credentials were steered toward premium AI models over general cloud access.
  • The Cloudflare case: Cloudflare’s responders saw ARTEX reach a compromised account-level API token used for CI/CD with no IP restrictions or rate limiting. What followed included 400 database dump requests in an hour; the company has not said whether they returned data, nor how the token was stolen.
  • The footprint: New Worker scripts named artex-*, artex-pgrecon-*, artex-pgw-* and artex-do-* appeared in the account, each name matching a task, running at more than 69 requests per minute. Cloudflare believes a human operator lightly guided much of it, and has not disclosed whether persistence attempts succeeded.
  • The second campaign: Decoy records also describe agents trying to bypass identity checks at two cryptocurrency exchanges using deepfakes. One verification vendor’s server rejected the live deepfake video on every recording attempt. Amazon did not say whether the forged ID document photos the agents fabricated passed document checks.
  • The wider picture: The FBI said in a FLASH alert this week that a range of attackers have used ARTEX since at least August worldwide, summarizing 11 earlier reports. Cloudflare counted more than 50 families of AI penetration testing frameworks scanning its customers over three days.

Background:

ARTEX splits an operator’s goal into tasks: a planner agent assigns worker agents that run security tools, sending their work to a language model to decide each next step. Its developer pulled the GitHub repository on October 8 after abuse reports, but forks and a commercial ARTEX PRO version keep it available.

Between the lines:

Amazon’s own framing undercuts the novelty: none of the techniques were new, and what agents change is attacker speed and persistence. The entry points in the playbook are default credentials, unfinished installs, unauthenticated endpoints and unpatched software — failures of basic hygiene, not exotic exploitation. The self-funding loop, where stolen keys run the models that find more keys, is what lets the campaign scale without the operators paying for compute.

What’s next

Watch whether Amazon or Cloudflare confirms outcomes the agents only claimed, and whether One API publishes affected versions or a fix for the isCliRequest bypass. RAGFlow’s advisory still lists no patched version for CVE-2026-28797.

What to read next