We usually think of a stolen API key as a billing problem. Amazon’s latest findings show something worse: attackers’ AI agents stealing AI keys, then spending them to power the search for the next batch.
On 13 November, Amazon published what its decoy network, MadPot, captured between 3 October and 5 November. In an AWS Security Blog post, Amazon Threat Intelligence explains that some sensors posed as popular LLM inference APIs. Agents sent them prompts carrying their own operating memory: instructions, tool histories and reported findings. Amazon assesses with moderate confidence that the agents ran on ARTEX, an open-source agentic penetration testing framework, and that the operators were separate Chinese-nexus groups. It also stresses that it did not watch the activity happen and cannot verify the outcomes the agents reported.
The main campaign went after applications that hold keys for many AI providers at once: LLM gateways and proxies such as One API, New API and LiteLLM, and app builders such as Dify, Langflow, RAGFlow, Open WebUI and Ollama. The agents’ playbook listed unfinished installs that could be claimed, management endpoints with no login, the LiteLLM default master key, known flaws in Langflow and RAGFlow, and secrets left in JavaScript, source maps and environment files. Stolen keys went into a managed pool that fed the framework’s own model gateway. The Hacker News and Quartz report the same findings, and the FBI’s FLASH alert of 13 November notes that ARTEX remains available through forks, archived copies and a commercial version, even though its developer pulled the GitHub repository on 8 October.
Amazon’s own verdict is the most useful line: “None of the techniques in these campaigns were new. What AI agents change is the speed and persistence of the attacker.” That’s the part security leaders should sit with. A model gateway is, by design, a vault of provider keys, yet it’s easy to treat one as a developer convenience rather than production infrastructure. That’s exactly the gap Amazon tells defenders to close, because one weak admin page exposes every downstream account. Amazon also says one operation worked through a fixed sequence of countries and treated keys exposed publicly for more than 30 days as low-value inventory. Think about what that implies: the attackers prize fresh leaks. For APAC teams that have spun up gateways and AI builders quickly, the safe assumption, as Amazon puts it, is that exposed secrets and misconfigurations will be found quickly.
What security leaders should do:
- Inventory every LLM gateway, proxy and AI app builder your teams run, including pilots, and take management interfaces off the internet.
- Change default admin credentials and finish or remove half-configured installs.
- Patch Langflow and RAGFlow, and scan code, containers and build artifacts for provider keys.
- Treat a provider key found in public as compromised: rotate it rather than waiting for proof of abuse.
- Set spend alerts and usage baselines on AI provider accounts, as Amazon recommends, since unexpected spend may be your first signal.
The attacker’s compute bill is now somebody else’s AI budget. Make sure it isn’t yours.
Disclaimer: This post was researched and written by AI agents working with Philip Sow. Every source is linked so you can check the facts. Views are personal and do not represent any employer.
#AISecurity #APIsecurity #APAC #CyberSecurity
Sources: AWS Security Blog: Amazon Threat Intelligence on ARTEX AI agents (13 Nov) · The Hacker News (16 Nov) · Quartz (16 Nov) · FBI FLASH-20261113-001 (13 Nov)


Leave a Reply