⌁AI·CYBER·BRIEF▌
AI Threats7 min read

Amazon: AI agents raided exposed LiteLLM, Dify and Ollama servers for API keys; FBI issues ARTEX alert

Amazon's decoys captured ARTEX agents stealing AI provider keys from exposed gateways and trying deepfake KYC bypasses. Lock down self-hosted AI tools and rotate keys.

Amazon says AI agents built on the open-source ARTEX framework hunted for internet-exposed AI tools such as LiteLLM, Dify, Langflow and Ollama to steal the paid model API keys stored inside them, and separately tried to beat crypto-exchange identity checks with deepfakes. On the same day, November 13, 2026, the FBI published its own alert on ARTEX. If your organization runs a self-hosted AI gateway or app builder, check that it is not reachable from the internet without a login, and rotate the provider keys it holds.

What happened, in plain English

Many companies now run their own "AI front desk": a gateway or app-building tool (LiteLLM, One API, Dify, Langflow, Open WebUI, Ollama and similar) that sits between staff and paid AI services such as OpenAI or Anthropic. To work, these tools store the company's API keys (application programming interface keys: long passwords that let software use a paid service and bill the owner).

On November 13, 2026, Amazon Threat Intelligence reported that attackers had set AI agents loose to find these tools on the internet, pull out the stored keys, test them, and keep the working ones in a pool. The stolen keys then paid for the agents' own AI usage, so victims funded the next round of searching. Think of a burglar who steals your fuel card and uses it to drive to the next house.

Amazon saw this from an unusual angle. Its MadPot decoy network included sensors that pretended to be popular AI model services. Some of the attackers' agents used them as their "brain", and because agent software sends the full conversation history with every request, the decoys captured the campaign's working memory between October 3 and November 5, 2026.

A second campaign targeted two cryptocurrency exchanges. Agents tried to pass "know your customer" (KYC) identity checks using real-time deepfake video and forged ID photos. According to Amazon, the exchange's verification vendor rejected the deepfake in every attempt the agents recorded.

Separately, the FBI issued FLASH-20261113-001, a public overview of how criminals use ARTEX, an open-source "autonomous penetration testing" framework.

Are you affected? What to do now

You should care if your organization runs any self-hosted AI gateway, proxy or app builder, or if developers have spun one up on a cloud server "just to try it". Amazon names One API, New API, Veloera, Sub2API, LiteLLM, Dify, FastGPT, Langflow, RAGFlow, Open WebUI, LobeChat and Ollama as targets. If you only use AI through vendor web apps such as ChatGPT or Copilot with single sign-on, there is little to do beyond the general hygiene below.

Check exposure

  • Inventory every AI gateway, app builder and model server, including test instances on personal cloud accounts. Ask development teams directly.
  • Confirm none of their admin consoles or APIs is reachable from the internet without authentication, and that default or hard-coded passwords have been changed.
  • Look at AI provider billing and usage dashboards for unexpected spend, new regions, unusual models or traffic at odd hours. Amazon specifically recommends monitoring for unexpected AI spend.
  • Scan code repositories, container images, build logs and paste sites for leaked provider keys. Amazon says agents searched GitHub, package registries, container images, paste sites and Telegram.

Fix, in order

  1. Put AI gateways and inference endpoints behind your normal access controls (VPN, zero-trust proxy or at minimum strong authentication). Amazon's advice is to treat them as production systems.
  2. Patch these tools. Amazon says the agents used known vulnerabilities, default credentials and exposed secrets, not new tricks.
  3. Rotate every upstream provider key stored in an exposed or unpatched instance, and set spending limits with your AI providers.
  4. Check firewall and proxy logs against the IP addresses Amazon published (in its blog post) and those in the FBI FLASH. Both are historical cloud and proxy addresses, so treat a hit as a lead, not proof. The FBI explicitly asks defenders to vet indicators before blocking.
  5. If you run identity verification (finance, crypto, onboarding), ask your KYC vendor how it detects injected or synthetic camera feeds, and keep server-side liveness checks.

For Microsoft 365 and ConnectWise customers: the FBI FLASH repeats guidance from the ConnectWise supply-chain campaign: review affected tenants by application ID, audit credential additions and consent grants, and preserve logs before removing anything.

The expert view

Technically, almost nothing in the Amazon report is new. Harvesting keys from exposed LLM proxies is classic "LLMjacking", which we have covered repeatedly, from Symantec's agent caught running on other people's models to the Railway breach. The entry points are the usual ones: unauthenticated consoles, default passwords, unpatched software and secrets in public places.

Three things are worth noting.

First, the economics loop. Stolen keys were fed into the framework's own model gateway and paid for further searching. That makes the campaign largely self-financing. The cost of attack falls to near zero for the operator and is billed to victims.

Second, the operating rules. The captured memory shows a managed process, not a hobbyist script. Operators set country order, prioritized enterprise production sites and graded each credential: L1 a lead, L2 a confirmed working API call, L3 a quantified impact. Agents were told to treat credentials exposed publicly for more than 30 days as "low-value inventory", on the assumption that someone else had already used or revoked them. The lesson for defenders is that the dangerous window is the first days after a leak, so rotation needs to be fast. The architecture was multi-agent: planner, worker, verification and compression agents sharing a persistent knowledge graph.

Third, the deepfake result is a useful reality check. The agents built a pipeline that ran at 11.5 frames per second on one consumer graphics card and positioned a face within 7 pixels of the target. Yet every recorded server-side verdict was a rejection. Current liveness checks held here. This is a data point, not a guarantee: Amazon only saw what the agents logged, and attackers iterate.

Attribution. Amazon assesses with moderate confidence that ARTEX was used, and with moderate confidence that the activity came from "disparate Chinese-nexus threat groups". The basis is simplified Chinese in prompts, China time-zone settings, targeting consistent with Chinese interests and tooling mainly used by China-nexus actors. Prompts reportedly excluded mainland China and government domains. That is Amazon's assessment; the FBI FLASH does not attribute these two campaigns.

The FBI's wider picture. The FLASH says malicious ARTEX use has been observed since at least August 2026 across finance, government, technology, defense, semiconductors, energy, healthcare, manufacturing, telecoms and legal. Some intrusions went from first targeting to privileged access in under three hours. The FBI counts more than 1,600 publicly exposed ARTEX consoles as of early November and says the real number of deployments is likely much higher. The original developer removed the GitHub repository in October, but forks, localized variants and a commercial "ARTEX PRO" keep it alive. ARTEX has no model of its own. The FBI says attacker-controlled instances mostly use Chinese models, with DeepSeek Flash-family models dominating because of cost, capability and a lack of observable guardrails when the request is framed as penetration testing. Most of the campaigns it summarizes, including Storm-3358 against financial firms, have been reported before; the FLASH is valuable mainly as a consolidated, official checklist.

What is still unknown. Amazon says it could not verify the outcomes the agents reported, so "hundreds of targets" describes intent and activity, not confirmed breaches. Neither source says how many keys were stolen or how much money victims lost. Amazon says it notified potentially affected AWS customers and does not say AWS itself was compromised.

The defensive takeaway is unglamorous. AI agents make attackers faster and cheaper, but in these cases they walked through doors that were already open. Closing exposed consoles, removing default credentials and rotating leaked keys quickly blunts most of what is described here.

Official sources

Get the daily brief

AI + security signal by email: headlines, a two-line summary, a link. No noise, no spam.

How often