Amazon caught AI agents pulling LLM keys from exposed gateways like LiteLLM and Dify, then spending them to find more
Amazon's threat team watched attacker agents steal AI provider keys from self-hosted tools and use those keys to pay for the next round of theft.
What happened
From October 3 to November 5, Amazon's MadPot decoy sensors posed as popular LLM inference APIs. Agents running on ARTEX, an open-source agentic penetration testing framework, sent requests to them, and those requests carried the agents' own campaign history: system prompts, tool histories and findings.
The largest campaign Amazon describes was credential theft for AI model access. The agents looked for internet-exposed AI applications, extracted the upstream provider keys they held, validated them and stored them in a pool. Stolen keys then powered the agents searching for more. Targets included LLM gateways such as LiteLLM and One API, and app builders such as Dify, Langflow, Open WebUI and Ollama. The playbook listed default admin credentials including the LiteLLM default master key, Langflow's CVE-2025-3248, anonymous management endpoints, and secrets left in client-side JavaScript, source maps and environment files. Keys exposed for more than 30 days were graded as low value, which suggests fresh leaks get used first.
Amazon notes it cannot verify the outcomes the agents reported. Its conclusion is that none of the techniques were new and the agents succeeded where basic hygiene was missing. Separately, TechCrunch reports an FBI advisory saying ARTEX-linked attacks can go from targeting to compromise in under three hours, and that Barracuda was breached through a third-party credential it never revoked after the vendor relationship ended.
My take
Small teams often run these exact tools. A LiteLLM proxy or an Open WebUI instance on a VPS, holding OpenAI and Anthropic keys for several workflows, is a common setup, and it is now a target that gets scanned automatically.
The fixes are ordinary and fast. Change default master keys and admin passwords on day one. Keep management interfaces off the public internet, behind a VPN or an IP allowlist. Patch builders like Langflow when advisories come out. Issue a separate provider key per client or per workflow so one leak has a known blast radius, and set spend limits and usage alerts at the provider, because unexpected spend is often the first visible sign. When a contractor or vendor engagement ends, rotate every key they touched.
More posts
- Open OCR models now read 1,000 pages for under $0.10, and better serving saved more than switching modelsNov 20, 2026
- Gemini 4 Neon costs $3 in and $15 out per million tokens and tops Zapier's AutomationBench at 56.3%, though it uses more output tokens per taskNov 20, 2026
- Gemini API's Tool Retrieval loads tool schemas only when the model asks for them, so agents with hundreds of tools send fewer input tokensNov 20, 2026
