Robots Are Better Than People home
Subscribe
Rows of server racks in a dim data center with a single glowing monitor showing network traffic logs

Amazon decoys catch attacker AI agents hunting AI keys to pay for their own hunt

Attackers pointed AI agents at exposed AI apps to steal provider keys, then used those keys to power more hunting. The agents gave themselves away by sending their own campaign notes to Amazon's decoy servers.

Amazon's threat intelligence team says AI agents run by attackers searched internet-exposed AI apps such as LiteLLM, One API and Dify for the AI provider keys stored inside them. The team learned this from the agents themselves. Between October 3 and November 5, 2026, the agents sent requests to Amazon decoy servers posing as popular large language model services, and those requests carried the campaigns' own history.

That matters for anyone running AI tooling on the open internet. According to Amazon, the stolen keys paid for the model inference that powered the search for more keys. Amazon also says the agents succeeded where basic security was missing, which means most of the fixes are within reach of small teams.

The attackers' agents reported to the defenders' decoys

Amazon runs MadPot, which it describes as a globally distributed network of decoy threat sensors. In this case the sensors presented themselves as LLM chat-completions interfaces. Agent frameworks treated them as real model backends and sent them prompts. Because these frameworks keep conversation history, a single request often held a campaign's full record: system prompts, tool histories, operator instructions and reported findings.

Amazon assesses with moderate confidence that the agents ran on ARTEX, an open-source agentic AI penetration testing framework. In ARTEX, an operator sets a target and a goal, a planner agent splits the goal into tasks, and worker agents run security tools, sending their work to an LLM to decide each next step. Amazon also assesses with moderate confidence that the operators were separate China-linked groups, citing Simplified Chinese prompts and China time zone settings, among other evidence.

This is the part we find most interesting from a defender's seat. A decoy designed to look like an AI service caught attackers whose tooling depends on AI services, and the attackers' own automation handed over the notes.

The targets are the AI plumbing many small teams run

The records name LLM gateways and proxies (One API, New API, Veloera, Sub2API, LiteLLM) and AI app builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama). The agents favored apps that hold keys for many providers, because one weak admin page there can expose many accounts. The records cover hundreds of targets, and Amazon says it notified AWS customers who may have been affected.

The agents' playbook leaned on familiar weaknesses: default administrator credentials such as LiteLLM's default master key, unfinished installs an outsider could take over, settings endpoints with no login, and known product flaws. When agents found cloud credentials, they went after managed AI model services, secret stores and application identities, prioritizing premium AI models over general cloud access. Here is the patch status for the flaws named, per The Hacker News:

ProductFlawStatus
LangflowCVE-2025-3248, missing authentication allowing remote code executionFixed in 1.3.0. In CISA's KEV catalog since May 5, 2025, marked as known to be used in ransomware campaigns.
RAGFlowCVE-2026-28797, server-side template injection that lets any logged-in user run commandsNVD lists 0.24.0 and earlier as affected. The advisory lists no patched version; 0.25.0 renders templates in the recommended sandbox. Self-registration is on by default, per the advisory.
One API familyAuthentication bypassAmazon's report names no affected versions or fix.

A liveness check beat the deepfakes every time

The decoy records also describe a second campaign: agents trying to bypass identity checks at two cryptocurrency exchanges with real-time deepfake video and forged ID photos. According to The Register, ARTEX ran several agents in parallel toward the same goal. "One searched for ways to cash out, one worked on clearing account balances, one set up the deepfake video pipeline, and one carried out the verification bypass," Moses wrote.

Field note: The agents studied one verification vendor's liveness check in detail and completed its movement sequence, yet the vendor's server-side verdict rejected the deepfake on every attempt, Amazon reports. Amazon did not say whether the forged ID photos passed document checks.

The Messy Middle

Much of this story rests on the attackers' own claims. "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. Attribution to ARTEX and to China-linked groups is at moderate confidence, not certainty.

Other reports add weight without closing gaps. Cloudflare says it saw ARTEX use a compromised account-level API token, previously used for CI/CD, with no IP restrictions or rate limiting. The activity included 400 database dump requests in an hour. Cloudflare has not said whether those requests returned data, how the token was first stolen, or whether persistence attempts worked. Microsoft reported attackers using agentic tools such as ARTEX to hunt GitHub for secrets, including AI inference keys.

The FBI's FLASH alert says attackers have used ARTEX since at least August, but it summarizes 11 earlier reports and describes no new incident. The alert states that "ARTEX significantly reduces the time required to chain attack techniques, with some intrusions progressing from initial targeting to privileged access in under three hours." The developer pulled the GitHub repository on October 8, yet forks, archives and a commercial version called ARTEX PRO remain. Censys counted 1,465 hosts running an exposed ARTEX console as of November 4, per The Register.

Most fixes are basic hygiene, applied fast

"None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker," Amazon's team said. Moses put it plainly: "Getting those fundamentals right remains the most effective defense."

  • If you run LiteLLM or another gateway: change default admin credentials, including LiteLLM's default master key, and never leave a fresh install unfinished on the internet.
  • If you run Langflow or RAGFlow: upgrade Langflow to 1.3.0 or later, and review RAGFlow's advisory and 0.25.0 notes.
  • If you store keys in code or pipelines: scan repositories, container images and config files for secrets and rotate exposed ones quickly, as the FBI advises.
  • If you use Cloudflare API tokens: set expiry dates and client IP limits, and look for Worker scripts named artex-*, artex-pgrecon-*, artex-pgw-* or artex-do-*.
  • If you run detection: check logs against the 11 IP addresses Amazon published and the more than 2,500 addresses Cisco Talos links to ARTEX.

The Inflection Point

Amazon says the agent records show these frameworks have moved from experimental use to sustained, multi-objective operations running with minimal human intervention. The key-theft loop is the notable shift: stolen AI access funds more automated attacks, so an exposed key is both a loss and fuel. Cloudflare says more than 50 families of AI penetration testing frameworks tried to scan its customers over three days. Amazon's advice is to treat AI apps, model gateways and inference endpoints like any production system, assume exposed secrets will be found quickly, and watch for misuse through unexpected spending.

The bottom line: Attacker agents are fast, but in these records they won where defaults and leaked keys were left in place and lost to a well-built server-side check. Close the basics before the agents find them.

Sources

We checked every fact, figure and quote in this story against the sources below. Quotes are reproduced as published. Images are AI-generated illustrations, not photos of the events described.

  1. The Hacker News: Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys (2026-11-16)
  2. AWS Security Blog: Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents (2026-11-13)
  3. The Register: Attackers are using AI agents to automate cybercrime at 'industrial scale,' FBI and AWS warn (2026-11-16)

RABTP Newsroom

Every story is researched against primary sources, then reviewed by publisher Sean Hakes before it goes live. We cover threats from the defender's side and never publish exploit details. Read our standards or send a correction.

Keep reading

The White-Hat Brief

Weekly: what the machines learned, and how to stay ahead

One email a week. Verified stories, plain-language defenses, no hype. Unsubscribe anytime.