
Amazon says AI agents targeted LiteLLM, Dify and other AI apps across hundreds of systems
Amazon Threat Intelligence captured records from AI-agent campaigns that targeted exposed LiteLLM, Dify and other AI platforms to steal provider API keys, describing hundreds of targets and a detailed exploit playbook. The evidence shows autonomous agents can hunt, exploit and verify credentials across AI infrastructure, turning weak admin pages into multi-provider key theft.
Amazon Threat Intelligence said it captured agent records using decoy servers that posed as popular large language model services. The records describe AI agents searching internet-exposed AI applications, including LiteLLM, One API, New API, Dify, Langflow, RAGFlow, FastGPT, Open WebUI, LobeChat and Ollama, then extracting provider keys stored inside. Amazon said the agents verified working keys and added them to a shared pool, focusing on gateways that hold credentials for many providers because one weak admin page can expose many accounts.
The evidence comes from MadPot, a decoy network that received agent requests carrying campaign history, operator instructions and findings. Amazon said it cannot verify outcomes on systems it did not observe, but the records show what the actors pursued. The playbook listed default administrator credentials, unfinished installs, no-login management endpoints, an isCliRequest header bypass in One API family systems, SSRF through URL-fetch features to cloud metadata services, and secrets in JavaScript, source maps, debug routes and environment files.
Amazon assessed with moderate confidence that the operators were separate China-linked groups, citing Simplified Chinese prompts and China time zone settings. The campaign used an operator-supplied goal split into tasks by planner and worker agents that ran security tools and consulted an LLM for next steps. Amazon had moderate confidence the agents ran on ARTEX, an open-source penetration-testing tool. Cloudflare separately reported ARTEX access to a compromised CI/CD API token with no IP restrictions or rate limiting, followed by 400 database dump requests in an hour.
The records show that exposed AI platforms can become credential warehouses, not just isolated services. A single unpatched gateway or misconfigured builder can leak keys for multiple model providers, enabling abuse of premium AI services, cloud metadata access and downstream CI/CD systems. Amazon notified AWS customers who may have been affected. The operational significance is that defenders must treat AI app admin pages, default keys, client-side secrets and CI/CD tokens as high-value targets, because automated agents can search, exploit and verify access at machine speed.
Organizations running exposed AI gateways, app builders and CI/CD systems can leak many model-provider credentials through one weak admin page or token. Automated agents can search, exploit and verify that access at machine speed.
Sources (1)
- [1]Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys(https://thehackernews.com/2026/11/amazon-uncovers-ai-agent-campaign.html)