Just in
TechNvidia Earnings: Q3 Revenue Up 103% to $115.8 BillionTechApple Outage: iCloud Sign In Issue Continues as App Store ClearsTechARTEX AI Agents Used in Key Theft and Deepfake Attacks, Amazon SaysTechNvidia National Security Board Committee Will Oversee Export RulesTechOpera Earnings: Q3 Revenue Up 20% to $182.8 Million, Outlook RaisedAICalifornia Sues OpenAI, Says Its AI Agents Attacked Hugging Face
Tech

ARTEX AI Agents Used in Key Theft and Deepfake Attacks, Amazon Says

Amazon Threat Intelligence says groups it links to China ran ARTEX AI agents to steal AI service keys and to test deepfakes against identity checks at two crypto exchanges.

USA Publishing Group4 min read
Share
Close view of programming code on a dark monitor, the kind of exposed AI app code ARTEX agents searched for keys
Photo: Shahadat Rahman / Unsplash

Key takeaways

  • Amazon says threat groups it links to China, with moderate confidence, used the open source ARTEX framework to run AI agents against hundreds of targets.
  • One campaign hunted API keys for AI models through exposed gateways such as LiteLLM, Langflow and RAGFlow, while another built deepfakes to beat identity checks at two crypto exchanges.
  • Amazon notified affected AWS customers where appropriate, published 11 IP addresses as indicators and urged companies to remove default credentials and patch AI tools.
In this story 4 sections
  1. How Amazon read the agents' own notes
  2. Campaign one: stolen keys to run AI for free
  3. Campaign two: deepfakes aimed at crypto exchange checks
  4. What Amazon did and what it recommends

Attackers that Amazon links to China used ARTEX, an open source AI agent framework built for penetration testing, to hunt for stolen AI service keys and to try to fool cryptocurrency exchange identity checks with deepfake video, according to a new report from Amazon Threat Intelligence. Amazon (NASDAQ: AMZN) published the findings on Friday, Nov. 13, in an AWS Security Blog post signed by CJ Moses, chief information security officer of Amazon Integrated Security. Searches for ARTEX AI agents climbed on Google Trends in the US on Tuesday.

How Amazon read the agents' own notes

Amazon did not spot the operations by watching victims. Its MadPot network of decoy sensors posed as popular large language model chat APIs from Oct. 3 to Nov. 5. When ARTEX agents called those fake endpoints, they sent along their working memory: system prompts, task context, a history of the tools they had run, a shared knowledge graph and their findings. That gave researchers a view of how the work was divided among planner, worker, verification and compression agents.

The company attributes the activity, with moderate confidence, to separate threat groups with ties to China. It pointed to simplified Chinese in the prompts, China time zone settings, the choice of targets and the tooling. The agents were told to skip targets in mainland China and government domains and to send their traffic through proxies, the report said. Amazon said the agents went after hundreds of targets, but it also cautioned that it cannot verify the outcomes the agents reported.

Campaign one: stolen keys to run AI for free

The first campaign aimed to collect API keys and cloud credentials for AI models so the attackers could run those models at someone else's expense. Stolen keys were fed back into the framework's own model gateway to power more searching.

Targets were internet facing LLM gateways and AI app builders. Products named in the report include:

  • One API, New API, Veloera and Sub2API
  • LiteLLM, Dify, FastGPT and Langflow
  • RAGFlow, Open WebUI, LobeChat and Ollama

Methods included known flaws such as CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow, an authentication bypass in the One API family, server side request forgery aimed at cloud metadata services, and default credentials such as the LiteLLM default master key. The agents also combed GitHub, npm and PyPI packages, container image layers, notebooks, paste sites and Telegram channels for exposed secrets. Each find was graded on three levels: a lead, a key confirmed with a working API call, and a key with measured impact. Keys that had sat in public for more than 30 days were treated as low value, according to Amazon.

Campaign two: deepfakes aimed at crypto exchange checks

The second campaign tried to get past know your customer checks at two cryptocurrency exchanges, which Amazon did not name. Agents produced live deepfake video for selfie liveness tests and forged ID photos for document checks, and they edited JPEG metadata so altered images would pass as untouched camera shots.

One session reverse engineered a verification vendor's liveness protocol and, in its best attempt, placed a face within 7 pixels of the target. The vendor's server still rejected the deepfake in every recorded attempt, Amazon said. Another session assembled a live deepfake pipeline that ran at 11.5 frames per second on a single consumer graphics card. Other workstreams mapped out cash out routes and ways to empty account balances.

What Amazon did and what it recommends

Amazon said it notified AWS customers that appeared in the activity where appropriate. Because it could not confirm what happened at other organizations, it published 11 IPv4 addresses as indicators of compromise, five tied to the key theft campaign and six to the deepfake campaign, so teams running the affected software can check their logs.

Its advice centers on basics. Remove default and hardcoded credentials, keep keys out of code and public packages, require authentication on inference endpoints and admin panels, patch quickly, and hold AI apps and model gateways to the same controls as other production systems. Amazon also suggested watching AI service bills and usage for jumps that could point to stolen keys.

"None of the techniques in these campaigns were new," Moses wrote. "What AI agents change is the speed and persistence of the attacker."

The findings add to a run of recent accounts of AI agents used in attacks, including a report on AI run exploits with no human in the loop and California's lawsuit over AI agents and Hugging Face. More security coverage is on our tech page.

ARTEX AI Agents Behind Key Theft and Deepfakes