
Amazon and the FBI warned on Thursday, November 13, 2026, that criminals are using autonomous AI agents to steal API keys at industrial scale, turning the keys they steal into the fuel that powers the hunt for more. Amazon's threat intelligence team captured the agents' own records on its MadPot decoy servers, and the FBI issued a FLASH alert the same day. Together, the warnings describe a new kind of cybercrime: attacks that run largely on their own, probing hundreds of targets.
Amazon Threat Intelligence published its findings on November 13, in a post by chief information security officer CJ Moses. Its MadPot network of decoy servers poses as popular AI services, and between October 3 and November 5 it received requests from AI agents that often carried a campaign's full history. The team said it had moderate confidence the agents ran on ARTEX, an open-source AI tool built for penetration testing, and that the operators were separate China-linked groups, citing Simplified Chinese prompts and China time-zone settings.
The FBI issued its FLASH alert the same day, saying attackers have used ARTEX since at least August against organisations in many sectors. The alert summarised 11 earlier reports, including a campaign against South Korean financial institutions that stole personal information on about 67,000 people, and the ConnectWise compromise.
ARTEX works like a small team. A human operator gives it a target and a goal, a planner agent splits the goal into tasks, and worker agents run the security tools. To decide each next step, the agents send their work to a large language model. That loop is what makes the attacks different: the agents keep probing without a human watching every move.
The agents followed a playbook of known entry points, including default administrator credentials such as LiteLLM's default master key, unfinished installs an outsider could take over, and known flaws in Langflow and RAGFlow. Targets included gateways such as LiteLLM, One API and Sub2API, and app builders such as Dify, FastGPT, Langflow, RAGFlow and Ollama. To find leaked keys, the agents searched GitHub, npm and PyPI packages, container layers, notebooks, model-hosting spaces, paste sites and Telegram channels.
The most striking detail is the self-funding loop. When the agents found keys, they verified they still worked, added them to a shared pool, and then used them to run the AI models that powered the search for more keys. They focused on apps holding keys for many providers, where one weak page can expose many accounts.
The flaws in the playbook were specific. Langflow carried CVE-2025-3248, a missing-authentication flaw that lets a remote attacker run code in versions before 1.3.0; CISA added it to its Known Exploited Vulnerabilities catalogue in May 2025. RAGFlow carried CVE-2026-28797, a template injection flaw that lets a logged-in user run operating system commands. One API systems could be bypassed through the isCliRequest header, and server-side request forgery could reach cloud metadata services.
Amazon was careful about what the records prove. "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. Its bluntest line: "None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker."
Cloudflare described one ARTEX case in which attackers reached a customer's environment through a compromised CI/CD API token with no IP restrictions or rate limiting. New Worker scripts named artex-* appeared in the account, running at more than 69 requests per minute, including 400 database dump requests in an hour.
AI agents have already proved they can find real flaws: Google DeepMind said this month its agents found two weaknesses in legacy 1024-bit RSA. ARTEX itself is not going away. Its developer made its GitHub repository unavailable on October 8, after reports of abuse, but community forks, archived copies and a commercial version called ARTEX PRO keep it available.
API keys are money. A stolen key for a leading AI provider can be spent on computing power, resold, or used to run further attacks, and the victim may only discover it through an unexpected bill. Because the agents automate the whole cycle, a single exposed gateway can be drained in hours. The FBI's under-three-hours figure shows how little time defenders now have.
The campaigns also show how the AI boom has widened the attack surface. Companies rushed to install AI gateways, chatbots and app builders, often with default settings, leaving the keys to their most expensive services sitting inside them. Japan unveiled an emergency cybersecurity package this month after a wave of breaches. Amazon's advice: treat AI apps, gateways and inference endpoints like any production system, assume exposed secrets will be found quickly, and watch for misuse through unexpected spending.
Amazon said it has notified AWS customers who may have been affected. The immediate advice is concrete: change default administrator credentials such as LiteLLM's default master key, finish or remove unfinished installs, patch Langflow to 1.3.0 or later, update RAGFlow, rotate exposed credentials, and give API tokens expiry dates and IP restrictions. Defenders should also check logs for the 11 published IP addresses and for Worker scripts named artex-*.
Longer term, the warnings point to a lasting shift. With ARTEX forks and a commercial version still available, agent-driven attacks are likely to keep growing. The same agentic technology is now mainstream: Anthropic's latest models can coordinate more than 1,000 agents at once. Defenders should treat AI systems as prime targets, not as side projects.
What is ARTEX? ARTEX is an open-source, AI-powered penetration-testing framework. An operator gives it a target and a goal, and its agents use a large language model to run security tools.
What is an API key, and why would criminals steal one? An API key is a secret credential that lets software use a paid service, such as an AI model. A stolen key can be spent on computing power or resold.
Which apps were targeted? The records named gateways including LiteLLM, One API and Sub2API, and app builders including Dify, FastGPT, Langflow, RAGFlow and Ollama.
How do I know if my systems were affected? Look for unexpected AI spending, unusual usage and unknown API keys in use, and check for Worker scripts named artex-*.
What should I do now? Change default credentials, patch Langflow and RAGFlow, remove unfinished installs, rotate exposed keys, and set expiry dates and IP limits on tokens.