Amazon threat intelligence and the FBI have warned that attackers are using ARTEX, an open-source AI penetration-testing framework, to scan for exposed AI applications and steal the model-provider API keys stored inside them. Amazon published its findings on Monday, three days after the FBI released a FLASH advisory on the same tool.
How Amazon’s MadPot Decoys Captured ARTEX Agent Activity
Amazon captured records from attacker-run AI agents using MadPot, its network of decoy servers that posed as popular large language model services. The agents hunted for exposed AI apps and gateways, extracted the AI provider keys stored in them, verified that the keys worked, and pooled them. Those stolen keys then paid for the models that powered further hunting, creating a self-funding loop. MadPot began capturing the agent traffic in early October.
Amazon said it has moderate confidence that the agents ran on ARTEX and assessed, also with moderate confidence, that the operators were separate China-linked groups, citing Simplified Chinese prompts and China time-zone settings. The company said it cannot verify the outcomes the agents reported.
Which AI Gateways and App Builders the Agents Targeted
The captured records name LLM gateways and proxies including One API, New API, Veloera, Sub2API and LiteLLM. They also list AI app builders such as Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. Amazon described hundreds of targets appearing in the data.
The Playbook for Stealing Stored Provider Keys
The agent playbook starts with default admin credentials, such as LiteLLM’s default master key, and unfinished installations. It also tries unauthenticated management endpoints, an authentication bypass using the isCliRequest header in One API family systems, and server-side request forgery to reach cloud metadata services. Secrets left in client-side JavaScript, source maps and environment files are further sources.
Beyond live systems, the agents searched GitHub, npm, PyPI, container layers, notebooks, paste sites and Telegram for leaked keys. Amazon also described a second campaign that tried to defeat identity checks at two cryptocurrency exchanges with deepfakes. One verifier rejected the live deepfake video every time.
FBI FLASH Advisory Ties ARTEX to Privileged Access in Under Three Hours
The FBI’s 13-page FLASH advisory, released on Friday with indicators of compromise, says multiple actors have weaponized ARTEX since at least August. Some intrusions moved from initial targeting to privileged access in under three hours. The bureau describes ARTEX as a Go framework that plugs into external LLMs, including Claude, GPT, DeepSeek, GLM, Grok and Kimi, and divides goals between planner and worker agents.
The advisory also connects ARTEX use to a recent ConnectWise supply-chain breach that compromised Microsoft 365 customer environments. The affected sectors named in the reporting are finance, government, technology, defense and healthcare.
A Removed Repository That Keeps Circulating
The original GitHub repository, published by a developer using the handle “Autumn-27,” went live in July and was removed on October 8. Forks, archived copies, localized versions and a commercial variant called “ARTEX PRO” remain available. Researchers counted more than 1,600 exposed ARTEX consoles as of early November.
Cloudflare Reports ARTEX Activity Using a Stolen CI/CD Token
Cloudflare separately reported ARTEX activity that used a compromised account-level CI/CD API token. The token carried no IP restrictions or rate limits, and the activity included 400 database dump requests within an hour. The report does not state whether any data came back. Cloudflare identified static, non-expiring CI/CD tokens as the common root cause.
What the ARTEX Campaign Signals for Exposed AI Infrastructure
Amazon notified potentially affected AWS customers, and the malware researcher group MalwareMustDie released an open-source detection signature pack on the same day as the FBI advisory. Amazon said the low skill barrier and the speed of the agents are the main concern.
The pattern fits a trend in which AI infrastructure holds credentials that are directly monetizable. An AI provider key is valuable because it converts into compute at the victim’s expense, and in this campaign it also converts into more attack capacity. Removal of the original repository did not stop distribution, which suggests takedowns of open-source offensive tooling have limited effect once forks exist. The Cloudflare finding adds that the weakness is not confined to AI apps: long-lived tokens in build pipelines gave the same agents a fast route to databases.