AI Agents Are Hunting for Stolen API Keys in Exposed AI Apps, Amazon Finds
Amazon's threat intelligence team says attackers are using AI agents to search internet-exposed AI apps for the provider keys stored inside them. The evidence comes from the agents' own records, captured by decoy servers (part of a network called MadPot) that posed as popular LLM services. Amazon has moderate confidence the agents ran on ARTEX, an open-source AI penetration testing tool, and that the operators were separate China-linked groups. It also says it cannot verify the outcomes the agents reported.
The records include a playbook of entry points for each target. These cover default administrator credentials such as LiteLLM's default master key, unfinished installs an outsider could take over, and known flaws in Langflow and RAGFlow. They also list management endpoints that need no login, an authentication bypass in One API family systems, and server-side request forgery used to reach cloud metadata services. Secrets left in client-side JavaScript, debug routes and environment files are covered too. The agents verified that stolen keys worked and added them to a shared pool, then used them to power the AI models that searched for more keys.
The agents also hunted for leaked keys across GitHub, npm and PyPI packages, container image layers, paste sites and Telegram channels. Amazon notified AWS customers who may have been affected. Separately, Cloudflare's incident responders reported ARTEX activity in a compromised account-level API token used for CI/CD. The token had no IP restrictions or rate limiting, and the activity included 400 database dump requests in an hour. Cloudflare has not said whether any data was returned. The common root cause in several incidents was static CI/CD tokens that never expire.
Cloudflare also saw more than 50 families of AI penetration testing frameworks scanning its customers over three days.
Why it matters
Any business running a self-hosted AI gateway or app builder is exposed, because these tools concentrate keys for many providers behind one admin page, so a single weak default can expose many accounts. Static CI/CD tokens with no IP restrictions or expiry break the assumption that a stolen credential has limited use, since automated agents can test and exploit them at machine speed. Leaders must now treat AI provider keys as high-value secrets with the same rotation and exposure controls as cloud credentials.
What to do
- Change default credentials such as LiteLLM's default master key on every AI gateway and app builder.
- Restrict internet access to management and settings endpoints, and complete any unfinished installs.
- Rotate static CI/CD tokens and add expiry, IP restrictions and rate limits to API tokens.
- Look for unexpected Worker scripts named artex-* in your Cloudflare account and scan repositories and client-side code for exposed keys.