Security Digest · Nov 13, 2026 · AI agent security · AI agent tooling

ARTEX AI Agents Hunt Exposed LLM Gateways for API Keys

Amazon saw ARTEX agents raid exposed LiteLLM, One API, Dify, Langflow and RAGFlow installs for AI keys. Check exposure; act now on default credentials.

Our verdict

Check exposure

Nothing here is a new bug: automated agents are finding self-hosted AI gateways and app builders that are exposed with default keys, unfinished setups or old Langflow and RAGFlow versions, so the fix is basic hardening and key rotation.

Who must act: Anyone running LiteLLM, One API, New API, Dify, Langflow, RAGFlow, Open WebUI, Ollama or a similar AI gateway or builder that the internet can reach, and anyone whose AI provider keys have appeared in code, images, notebooks or front-end bundles.

Who can relax: Teams whose AI tooling is reachable only on private networks or behind SSO, with unique admin secrets, current versions and no provider keys in client-side code.

Affected
Internet-reachable AI gateways and app builders with default or example admin credentials, unfinished setup wizards or unauthenticated management endpoints
Langflow before 1.3.0 (CVE-2025-3248)
RAGFlow 0.24.0 and earlier (CVE-2026-28797)
One API family systems accepting the isCliRequest header bypass (versions not named by Amazon)
Fixed in
Langflow 1.3.0
RAGFlow 0.25.0 (adopts a sandboxed Jinja2 environment; the advisory itself lists no patched version)
No vendor fix named for the One API isCliRequest bypass as of Nov 16, 2026
CVSS
Not scored
EPSS
100.0% (percentile 100.0)
CISA KEV
Yes, listed as known exploited
Exploitation
Exploited in the wild
Severity
High

What to do

  1. Find every self-hosted AI gateway, proxy and app builder you run and confirm none answer from the public internet without SSO or a VPN in front.
  2. Replace default or tutorial admin secrets, such as a LiteLLM master key copied from an example, with a long random value, then restart.
  3. Upgrade Langflow to 1.3.0 or later and RAGFlow to 0.25.0 or later, and turn off RAGFlow self-registration if you do not need it.
  4. Finish or remove any half-completed install whose setup page is still reachable.
  5. Rotate any AI provider or cloud key that was stored in an exposed gateway or appeared in JavaScript bundles, source maps, .env files, images or notebooks.
  6. Set spend limits and alerts on AI provider accounts, and block outbound access to cloud metadata endpoints from AI app containers.

How to check if you are affected or compromised

  • Look for unexpected spend or new usage patterns on AI provider accounts, especially on premium models, since stolen keys were used to power the attackers' own agents.
  • Search gateway and reverse-proxy logs for Amazon's published IP addresses and for requests to admin, setup or provider-configuration endpoints from unknown sources.
  • Check the gateway for virtual keys, users, providers or model routes that nobody on your team created.
  • Run Cisco Talos's ARTEX indicator files and Snort rules against network logs; on Cloudflare, look for Worker scripts named artex-*.

What happened#

On November 13, Amazon Threat Intelligence described two criminal campaigns run with ARTEX, an open-source agentic penetration testing framework. Amazon's MadPot decoys posed as popular LLM inference APIs. From October 3 to November 5 they captured requests from attacker agents, and those requests included the agents' own campaign records. In the main campaign, agents searched for internet-exposed AI gateways and app builders, pulled out the AI provider keys stored in them, tested the keys and pooled the working ones. Those keys then paid for the next round of searching. Amazon attributes it, with moderate confidence, to several separate China-nexus groups. A second campaign used deepfakes against exchange identity checks.

The same day, the FBI's FLASH-20261113-001 said several actors have used ARTEX since at least August 2026 across many sectors. Some intrusions went from first targeting to privileged access in under three hours. The original repository was pulled on October 8, but forks and a commercial "ARTEX PRO" remain. Cisco Talos links more than 2,500 command-and-control IP addresses to ARTEX. Cloudflare says more than 50 AI pentest framework families scanned its customers in one three-day window. In one case it handled, ARTEX used a stolen, non-expiring account-level API token to make 400 database dump requests in an hour.

Who is affected#

Amazon names the targets: the LLM gateways One API, New API, Veloera, Sub2API and LiteLLM, and the app builders Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama. The agents' playbook had nothing exotic in it:

  • default administrator credentials, including what Amazon calls the LiteLLM default master key
  • unfinished installs that an outsider can claim
  • management and provider-settings endpoints that need no login
  • an isCliRequest header that bypasses authentication in One API family systems
  • SSRF through URL-fetch features to reach cloud metadata services
  • keys left in client-side JavaScript, source maps, debug routes and .env files
  • two known flaws: Langflow CVE-2025-3248, an unauthenticated code-execution bug that has been in CISA KEV since May 5, 2025 with known ransomware use and is fixed in 1.3.0, and RAGFlow CVE-2026-28797, a template injection that any logged-in user can turn into OS commands

The RAGFlow advisory lists 0.24.0 as affected and names no patched version. The v0.25.0 release of April 21 switches to Jinja2's sandboxed environment. Self-registration is on by default, so "logged-in user" can mean anyone.

Do you really need to act?#

Only if your AI tooling is reachable from the internet, but then yes. Amazon is blunt that "none of the techniques in these campaigns were new." The agents won where basics were missing. What changed is speed and persistence: an agent that never gets bored will find a forgotten Dify install or a gateway still using a tutorial key. One exposed gateway can hold keys for several providers plus cloud credentials.

The limits matter too. Amazon watched decoys, not victims, and says it cannot verify the outcomes the agents reported. No new CVE is involved. If your gateway sits behind SSO or a VPN with a unique admin secret, this campaign is not a reason to drop other work.

What to do now#

  1. List every AI gateway and builder you run, including demo instances. From outside your network, check whether each answers: curl -s -o /dev/null -w '%{http_code}\n' https://ai-gw.example.com/v1/models should return 401 or fail to connect, never 200.
  2. Check the admin secret. In LiteLLM the master key authorizes every management call and is also the Admin UI password. Generate a real one as the official quick start does, with openssl rand -hex 32, then restart. Do the same for root or admin accounts in One API, Dify, Open WebUI and similar tools.
  3. Check versions: pip show langflow ragflow litellm, or docker ps --format '{{.Image}}' for container tags. Move Langflow to 1.3.0 or later and RAGFlow to 0.25.0 or later, and turn off RAGFlow self-registration if you don't need it.
  4. Remove any reachable setup or first-run page, and keep Ollama on a private interface.
  5. Rotate every provider key the gateway held if it was ever exposed. Search your own build output for keys: grep -rnE 'sk-[A-Za-z0-9_-]{20,}' dist/ build/ .next/static/.
  6. Set spend caps and alerts with your AI providers, and block container access to 169.254.169.254 unless something genuinely needs it.

How to tell if you were hit#

The clearest sign is money: sudden premium-model usage in your AI provider billing. Then look inside the gateway for virtual keys, users, provider entries or model routes nobody on your team created. Search reverse-proxy logs for Amazon's published IP addresses and for anonymous requests to admin, setup or provider-configuration paths. Talos publishes indicator files and Snort rules (Snort 3 7:76–7:78, Snort 2 1:67494–1:67496). The POLICY rules 7:78 and 1:67496, which detect the "norma" harness, are off by default. On Cloudflare, look for Worker scripts named artex-*. On any match, treat every key the gateway held as stolen.

Timeline#

  • — RAGFlow publishes the CVE-2026-28797 advisory. v0.25.0, with sandboxed templates, follows on April 21.
  • — Amazon's decoys begin capturing ARTEX agent traffic. The window runs to November 5.
  • — The ARTEX developer removes the GitHub repository.
  • — Amazon, the FBI, Cisco Talos and Cloudflare publish their findings.
  • — The Hacker News and others report it.

Vulnerability records

  • CVE-2025-3248 · KEV 2025-05-05 · Langflow versions prior to 1.3.0 are susceptible to code injection in the /api/v1/validate/code endpoint.
  • CVE-2026-28797 · RAGFlow is an open-source RAG (Retrieval-Augmented Generation) engine.

Sources

Original sources first: vendor and project advisories, CVE and government records, the discoverers’ write-ups. Every claim above links to one of these. See how the Security Desk works.

Original and primary sources

  1. Amazon Threat Intelligence — Threat actors running cybercrime campaigns with ARTEX AI agents aws.amazon.com
  2. FBI — FLASH-20261113-001: Threat Overview of ARTEX Autonomous Penetration Testing Framework ic3.gov
  3. Cisco Talos — ARTEX and beyond: AI-enabled penetration testing platforms blog.talosintelligence.com
  4. Cisco Talos — IOC files, November 2026 github.com
  5. Cloudflare — Adapting to AI penetration testing orchestrators blog.cloudflare.com
  6. CISA — Known Exploited Vulnerabilities: CVE-2025-3248 (Langflow) cisa.gov
  7. RAGFlow — GHSA-vvwj-fvwh-4whx: SSTI leading to RCE in Text Processing component github.com
  8. RAGFlow — Release v0.25.0 github.com
  9. LiteLLM — Docker quick start docs.litellm.ai

Coverage and analysis

  1. The Hacker News — Amazon Uncovers AI Agent Campaign Targeting LiteLLM, Dify, and Other Apps for API Keys thehackernews.com