ARTEX-run AI agents: default creds and app flaws in LiteLLM, Dify, One API steal provider API keys
Severity: high · Patch status: no-fix-needed · Since 2026-11-13 · CVE-2025-3248, CVE-2026-28797
Amazon Threat Intelligence published (2026-11-13) records captured by its MadPot decoy servers, which posed as LLM services and received requests from AI agents between 3 October and 5 November 2026 that carried full campaign histories. Amazon assesses with moderate confidence that the agents ran on the open-source ARTEX pentest framework and that the operators were disparate China-nexus groups (simplified Chinese prompts, China time-zone settings, targeting); it says activity against other organizations could not be confirmed and that the records describe activity across hundreds of targets. The agents scanned for internet-exposed AI apps and LLM gateways (One API, New API, Veloera, Sub2API, LiteLLM, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama), tried default admin credentials such as LiteLLM's default master key, unfinished installs, an isCliRequest auth bypass in One API-family systems, SSRF through URL-fetch features and known Langflow and RAGFlow flaws (CVE-2025-3248, CVE-2026-28797), then extracted and tested the provider keys stored inside and added them to a shared pool. They also searched public code, packages, paste sites and Telegram for leaked keys. Cloudflare reports a related ARTEX case in which a stolen account-level CI/CD API token with no IP restriction was used. The FBI's FLASH-20261113-001 on ARTEX, released the same day, says actors have used it since at least August 2026. Amazon has notified AWS customers as appropriate. No vendor patch applies to the campaign. Remediation: rotate AI provider and cloud keys held by any internet-exposed AI app or gateway, take those apps off the public internet behind SSO, change every default admin credential and finish or remove unfinished installs, and replace static CI/CD tokens with short-lived, IP-restricted ones.
What to do
- Rotate and revoke all AI provider API keys and cloud credentials stored in internet-exposed AI apps and gateways
- Remove AI apps and LLM gateways from public internet exposure; place them behind VPN/SSO with MFA
- Change all default admin credentials (e.g. LiteLLM master key) and complete or tear down unfinished installs
- Replace static, non-expiring CI/CD and account-level API tokens with short-lived, IP-restricted, rate-limited credentials
- Audit logs for ARTEX-style indicators (artex-* scripts, high-rate database dump or scanning requests) and SSRF attempts against cloud metadata endpoints
- Scan repos, packages, containers and paste sites for leaked keys and rotate any found
Sources
All live incidents on Threat Watch