Amazon Threat Intelligence has published a detailed analysis of an active campaign in which AI agents autonomously hunt for internet-exposed instances of LiteLLM, Dify, One API and related platforms to steal stored AI provider keys.
The agents, believed to be run by separate China-linked operators, leverage the stolen keys to execute the very large language models that drive their own reconnaissance and exploitation loops. Amazon captured hundreds of agent interaction logs through its MadPot network of decoy LLM endpoints between 3 October and late last week.
Attack Playbook and Targeted Products
The recovered records contain a structured playbook listing known weaknesses for each product family. Entries include LiteLLM’s default master key, unfinished installations left publicly reachable, the isCliRequest header authentication bypass affecting the One API family, and server-side request forgery vectors that reach cloud metadata services.
Additional techniques cover secrets left in client-side JavaScript, source maps, debug routes and environment files. The agents also scan certificate transparency logs, GitHub, npm, PyPI, container layers and paste sites for leaked credentials.
Known Vulnerabilities Exploited
- Langflow – CVE-2025-3248 (missing authentication allowing remote code execution); versions before 1.3.0 are affected and the flaw is listed in CISA’s KEV catalog.
- RAGFlow – CVE-2026-28797 (server-side template injection); affects 0.24.0 and earlier; version 0.25.0 introduces sandboxed template rendering.
Cloudflare incident responders separately observed ARTEX agents abusing an unrestricted CI/CD API token to issue more than 400 database-dump requests per hour and to deploy Worker scripts named artex-*, artex-pgrecon-* and artex-pgw-*.
Recommendations for Defenders
Amazon and Cloudflare urge organisations to treat AI gateways and inference endpoints as production systems: rotate default credentials immediately, enforce token expiry and source-IP restrictions on Cloudflare API tokens, and monitor logs for the 11 IPs published with the decoy findings as well as the more than 2,500 addresses linked to ARTEX by Cisco Talos.


