Flitz.ai Flitz.ai

November 16, 2026

Criminal AI Agents Are Hunting Self-Hosted AI Tools for API Keys — Check Yours Today

If your business has stood up a self-hosted AI chatbot, internal assistant or LLM gateway in the past year, there is new reason to check it right now. Amazon's threat intelligence team has confirmed t...

Criminal AI agents hunting self-hosted AI tools for API keys

If your business has stood up a self-hosted AI chatbot, internal assistant or LLM gateway in the past year, there is new reason to check it right now. Amazon's threat intelligence team has confirmed that autonomous AI "agents" are being used by criminals to find and raid exactly these tools — and the stolen credentials are being used to pay for even more attacks.

What actually happened

Between 3 October and 5 November 2026, Amazon's MadPot network of decoy sensors captured requests from AI agents that revealed their own operating memory from offensive campaigns. According to Amazon, these campaigns ran on ARTEX, an open-source agentic AI penetration testing framework repurposed for crime (Amazon Security Blog).

The named targets include widely used LLM gateways and proxies such as One API, New API, Veloera, Sub2API and LiteLLM, plus AI app builders including Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama (The Hacker News). These are exactly the tools many Swiss SMEs are now piloting to build their own internal chatbots and automation workflows.

The most striking part: the most prominent campaign was self-funding. Stolen API keys were used to pay for the very AI inference that powered further scanning and credential theft — a criminal operation that essentially bankrolls its own expansion (tbreak.com). Amazon assesses with moderate confidence that separate Chinese-nexus threat groups are behind these operations.

How they get in

None of this relies on sophisticated zero-days alone. The weaknesses exploited are depressingly familiar: exposed admin pages, open provider settings, and default passwords — including LiteLLM's default master key, which many deployments never change. Attackers also chained in known flaws like CVE-2025-3248 in Langflow and CVE-2026-28797 in RAGFlow (Security Affairs).

Separately, LiteLLM itself just patched two serious flaws in version 1.83.9: CVE-2026-35029, a missing-authorisation bug that let low-privilege users reach admin functions, and CVE-2026-59823, an SSRF flaw that let attackers redirect gateway traffic and steal AI provider keys (Check Point Research).

Why this matters for Swiss SMEs

If your business runs — or a vendor runs on your behalf — a self-hosted AI gateway or chatbot builder, an exposed instance is effectively an open wallet. Attackers don't need to breach your bank account; they just need your AI provider keys, which can rack up usage charges and be resold or reused to power further attacks against other targets, potentially implicating your business in someone else's crime wave.

What to do today

  • Inventory your AI tools. List every self-hosted LLM gateway, chatbot builder or AI proxy running in your business, including anything IT or a department set up as a pilot or proof of concept.
  • Change every default password and master key. LiteLLM's default master key is a known, actively exploited weak point.
  • Lock down admin pages. Make sure admin interfaces and provider settings are not reachable from the open internet — put them behind a VPN or allow-list.
  • Patch immediately. Update LiteLLM to version 1.83.9 or later. Check Langflow and RAGFlow instances against CVE-2025-3248 and CVE-2026-28797.
  • Rotate API keys for every AI provider connected to a self-hosted tool, and check provider billing dashboards for unexpected usage spikes.
  • Review access logs on any internet-facing AI tool for scanning patterns or unfamiliar requests over the past six weeks.

The bigger picture

This is part of a pattern we keep seeing: as businesses adopt AI tools faster than they secure them, criminals adopt AI tools to attack faster than defenders can patch. Self-hosted AI is genuinely useful for automation — but only if it's configured and maintained with the same discipline as any other business-critical system.

We'll keep watching this space and reporting on what matters most for Swiss SMEs. If you're running — or planning to run — self-hosted AI tools and want a second set of eyes on the configuration, Flitz.ai's team at The Mejlerö Company can help you automate your business securely, with a free network analysis and a free one-hour consultation to start.

Limited spots available this week

Get a FREE consultation

Get a FREE 1 hour consultation about your current application management setup, development process or any other tech setup problem.