Skip to content
NLEN
Illustration: OpenAI pipeline found 948 Linux root bugs in one run: what autonomous bug hunting means

AWS sees attackers using AI agents to hunt for API keys

By Ivo Donker · — text: Claude (AI)

What did Amazon see? Between October 3 and November 5, 2026, Amazon's MadPot decoy network captured requests from attackers using the open-source agent framework ARTEX. The agents searched at scale for API keys to AI models in poorly secured installations of LiteLLM, Dify, Open WebUI, Ollama and similar tools. In a second campaign they tried to defeat the identity checks of two crypto exchanges with deepfakes.

We read the Amazon Threat Intelligence blog post of November 13, written by CJ Moses, CISO of Amazon Integrated Security. The topic appeared earlier in the radar of November 17.

What happened

MadPot is Amazon's globally distributed network of decoy sensors. Some of them posed as a common chat-completions API for language models. Agents that used such an API for their own reasoning sent their requests to those decoys, and with them their own campaign records: which targets they looked for, what they found and how they graded it.

According to Amazon, ARTEX is an open-source framework for agent-driven penetration testing, with planner, worker, verification and compression roles that share a persistent knowledge graph. Amazon attributes the activity with moderate confidence to several China-nexus groups, based on simplified Chinese prompts, time-zone settings and tooling. The operators' instructions excluded targets in mainland China and government domains.

Details

Details
ItemAccording to Amazon
PeriodOctober 3 to November 5, 2026
Campaign 1Theft of API keys for AI models
TargetsGateways such as One API, New API, Veloera, Sub2API and LiteLLM; builders such as Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama
Entry pointsUnclaimed installs, anonymous admin endpoints, default credentials (including LiteLLM's default master key), CVE-2025-3248 (Langflow), CVE-2026-28797 (RAGFlow), SSRF to cloud metadata, secrets in JavaScript and environment files
Campaign 2Deepfake video and forged ID photos against the identity checks of two crypto exchanges
Outcome of campaign 2The verification vendor's server rejected the deepfake in every recorded attempt

Amazon says the stolen keys powered the framework's own model gateway, which in turn drove more key hunting. The agents searched internet scans, certificate transparency logs, GitHub, npm and PyPI packages, container images, notebooks, paste sites and Telegram. Each key was graded as a lead, a confirmed working API call or a quantified impact. Keys that had been public for more than 30 days were treated as low value. In the second campaign one agent built a real-time deepfake pipeline on a single consumer GPU that reached 11.5 frames per second.

What this means

Amazon puts it plainly: none of the techniques are new. What AI agents change is the speed and persistence of the attacker. A LiteLLM or Ollama server left open for an afternoon with a default password will be found quickly. Amazon advises assuming that exposed secrets and misconfigurations will be discovered fast, giving AI applications and model gateways the same controls as other production systems, watching for unexpected spending, removing default credentials, keeping keys out of code, requiring authentication and patching.

If you run your own models or gateway, see our overview of AI and cybersecurity and our article on a leaked API key at Cursor. On the deepfake side, read our article on deepfakes, and on how labs detect abuse, the article on misuse detection at OpenAI and Anthropic.

What we did not verify

Amazon did not observe the attacks directly and cannot verify whether what the agents reported in their records actually succeeded. The attribution has moderate confidence. We did not copy the IP addresses Amazon shares as indicators; they are in the blog post. We could not open a Google report on similar activity.

Checked on November 18, 2026.

This article was written with AI (Claude) and checked against the sources below.

Sources