NEWS

Amazon: AI Agents Are Hunting LiteLLM, Dify and Other AI Apps for Provider API Keys

Amazon decoy servers captured AI agents, likely running ARTEX, scanning exposed LLM gateways and app builders to steal stored AI provider keys.

Dylan H.

News Desk

November 16, 2026
4 min read
Amazon: AI Agents Are Hunting LiteLLM, Dify and Other AI Apps for Provider API Keys

Summary

Amazon's threat intelligence team used MadPot decoy servers to capture the working records of attacker-run AI agents. The agents were scanning internet-exposed AI applications to steal the AI provider API keys stored inside them. Amazon has moderate confidence the agents ran on ARTEX, an open-source AI penetration-testing framework, and assesses with moderate confidence that the operators were separate China-linked groups. Amazon could not verify the outcomes the agents reported.

This is separate from the FBI's flash alert on the same tool. See our earlier coverage: FBI Flash Alert: Artex AI Tool Takes Intrusions From Targeting to Privileged Access in Under Three Hours.

How the Campaign Works

  • An operator sets a target and goal. A planner agent splits the goal into tasks for worker agents that run security tools, and an LLM decides each next step.
  • Agents find exposed AI apps, extract stored provider keys, verify them and add them to a shared pool. Apps holding keys for many providers are prioritized.
  • Stolen keys then power the models that drive the search itself.

Techniques Observed

  • Default admin credentials, such as LiteLLM's default master key
  • Unfinished installs and unauthenticated management endpoints
  • An authentication bypass using the isCliRequest header in One API-family systems (no affected versions or fix were named)
  • SSRF to reach cloud metadata services
  • Secrets hunted in client-side JavaScript, source maps, debug routes and environment files
  • Leaked keys sourced from internet scanning, certificate transparency logs, GitHub, npm and PyPI packages, container images, notebooks, paste sites and Telegram channels

Targeted Applications

CategoryApplications
LLM gateways and proxiesOne API, New API, Veloera, Sub2API, LiteLLM
AI app builders and platformsDify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat, Ollama

Flaws named in the agents' playbook include:

  • Langflow CVE-2025-3248: missing authentication leading to remote code execution; fixed in 1.3.0; on CISA's KEV catalog and flagged as used in ransomware campaigns
  • RAGFlow CVE-2026-28797: server-side template injection letting any logged-in user run OS commands; self-registration is on by default; 0.24.0 and earlier are listed as affected, and 0.25.0 includes sandboxed template rendering

Cloudflare's Findings

Cloudflare reported ARTEX activity using a compromised account-level API token tied to CI/CD. The token had no IP restriction or rate limiting, and activity included 400 database dump requests in one hour. Cloudflare has not said whether any data was returned. New Worker scripts named artex-*, artex-pgrecon-*, artex-pgw-* and artex-do-* appeared, running at more than 69 requests per minute. The common root cause across incidents was static CI/CD tokens that never expired. Over three days, more than 50 families of AI pentesting frameworks scanned Cloudflare customers.

Context

  • The FBI says multiple attackers have used ARTEX since at least August across many sectors.
  • ARTEX's developer removed its GitHub repository on October 8, but forks, archived copies and a commercial version, ARTEX PRO, remain available.
  • Cisco Talos links more than 2,500 IP addresses to ARTEX. Amazon published 11 indicator IPs.
  • A separate campaign tried to bypass identity checks at two crypto exchanges using deepfakes; one verification vendor rejected the live deepfake video every time.

Recommendations

  1. Treat AI apps, model gateways and inference endpoints as production systems: no public admin pages, no default credentials, no unfinished installs.
  2. Rotate any provider key that has been exposed or stored in an internet-facing app, and set spend limits and alerts on each key.
  3. Scan configs, environment variables, CI/CD pipelines, repositories and container images for embedded secrets.
  4. Set expiry dates and client IP restrictions on Cloudflare API tokens, and look for artex-* Worker names.
  5. Patch Langflow (1.3.0 or later) and RAGFlow (0.25.0 or later).
  6. Check logs for the published IP indicators and consider the Cisco Talos Snort rules (7:76 to 7:78 for Snort 3; 1:67494 to 1:67496 for Snort 2). Rules 7:78 and 1:67496 are off by default.

Sources