Security
Amazon says decoy servers captured records of AI agents hunting provider keys in LiteLLM, Dify and One API apps, with moderate confidence that they ran on ARTEX
Amazon says decoy servers captured records of AI agents hunting API keys in exposed LiteLLM, Dify and One API apps. It did not see the activity live.
Amazon Threat Intelligence published findings this week on attackers using AI agents to hunt for AI-provider API keys in internet-exposed apps such as LiteLLM, One API and Dify. The evidence is records the agents themselves wrote, captured on Amazon’s decoy servers. Amazon says it has moderate confidence the agents ran on ARTEX, an open-source agentic penetration-testing framework. It also says it did not observe the activity as it happened and cannot verify the outcomes the agents reported.
The essentials
- What was captured: the agents’ own records, collected on decoy servers, not a live observation of the attacks.
- Targets: LLM gateways and proxies such as One API, New API, Veloera, Sub2API and LiteLLM, and app builders such as Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama.
- The loop: per the records, extracted keys were validated, pooled, and fed into the framework’s own model gateway, where they powered the agents searching for more keys.
- Limits: Amazon says outcomes are unverified; the records show what the actors were pursuing, not what they achieved.
Why gateways are attractive targets
Apps that store keys for many AI providers concentrate value: one weak admin page can expose many downstream accounts. The agents favored such apps, according to the records, and also treated cloud credentials as a route to more access.
What operators should take from it
If you run LiteLLM, One API, Dify or similar tools on the open internet, treat stored provider keys as exposed-by-default: keep admin interfaces off the public internet, rotate keys, set spend limits and patch. Amazon’s reported claims are unverified, but the targeting list is a useful checklist.
What is not established
- Whether any key was actually stolen: not verified, per Amazon.
- The number of victims or the amounts charged: not given in what we read.
- Who operated the agents: not named in the sources we read directly.
- Any link to a specific country: not established in what we read.
Summary table
| Point | What is stated as of November 7, 2026 | Source |
|---|---|---|
| Evidence | Agents’ own records captured on Amazon decoy servers | AWS Security Blog |
| Confidence | Moderate that the agents ran on ARTEX | The Hacker News |
| Verification | Amazon cannot verify outcomes the agents reported | The Hacker News |
| Targets | One API, New API, Veloera, Sub2API, LiteLLM, Dify, FastGPT, Langflow, RAGFlow and others | AWS Security Blog |
| Self-funding loop | Stolen keys fed the framework’s model gateway | Security Affairs |
FAQ
Did Amazon watch the attacks happen?
No. It says it did not observe the activity live and relies on records captured by decoys.
Which apps were targeted?
LLM gateways such as LiteLLM and One API, and AI app builders such as Dify, Langflow and Open WebUI.
Is ARTEX malware?
ARTEX is described as an open-source penetration-testing framework that attackers can also use.
Method and limits
Checked on November 16, 2026. We read AWS’s post, The Hacker News and Security Affairs through search excerpts, not full pages. The user-facing summary that agents ‘stole’ keys overstates what Amazon verified; we report it as claimed in the agents’ records.
Sources
- AWS Security Blog: Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents (verified 2026-11-08)
- The Hacker News: Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys (verified 2026-11-08)
- Security Affairs: ARTEX: How Cybercriminals Use AI Agents to Scale Their Attacks (verified 2026-11-08)
Spotted an error in this article? Report it · Our corrections policy