Title: Amazon Decoy Servers Capture AI Agents Self-Funding API-Key Theft Campaigns Assessed With Moderate Confidence to Have Run on ARTEX
Summary: Amazon Threat Intelligence reports that its MadPot decoy network, configured to pose as popular LLM inference APIs, captured requests between October 3 and November 5, 2026 from AI agents whose messages included the agents’ own records of offensive campaigns that Amazon assesses with moderate confidence ran on ARTEX, an open-source agentic AI penetration-testing framework. The headline campaign was large-scale theft of credentials for AI models, organized to be self-sustaining: agents found internet-exposed AI applications, extracted the upstream provider keys inside, validated them, and pooled them — with stolen keys then fed back into the framework’s own model gateway to pay for the inference powering further hunting. The targets fell into two groups: key-aggregating gateways and proxies (One API, New API, Veloera, Sub2API, LiteLLM) and application builders (Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama). A written playbook tied each fingerprinted product to its usual attack path, running from vendor default administrator credentials such as LiteLLM’s master key, half-initialized installs that anyone could take over, and management endpoints reachable without a login, through the Langflow flaw CVE-2025-3248 and the RAGFlow flaw CVE-2026-28797 and the isCliRequest-header authentication bypass in One API-family software, to request forgery abused in URL-fetch features to hit cloud metadata services, and credentials that had simply been left lying around in browser-side JavaScript, source maps, debug routes, or .env files. A second campaign used real-time deepfake video and forged ID document photos to defeat KYC liveness checks at two cryptocurrency exchanges; in every recorded attempt against one vendor, the server-side verdict rejected the deepfake. Amazon assesses with moderate confidence that these were separate China-nexus groups, citing Simplified Chinese prompts and China time-zone settings, and stresses it could not verify the outcomes the agents reported since the underlying activity ran outside MadPot’s view — but the records show what the actors were pursuing, and Amazon notified potentially affected AWS customers. The company’s takeaway for defenders: none of the techniques were new, and the agents succeeded wherever basic hygiene was missing — default or hardcoded credentials, keys in code and public artifacts, unauthenticated interfaces and inference endpoints, unpatched software — so organizations should treat AI apps, model gateways and inference endpoints like any other production system, assume exposed secrets will be found quickly, and watch for misuse through unexpected spending or unusual usage.
References: